Efficient privacy computing fusion engine method and system for medical data

Abstract

The application provides a kind of medical data-oriented efficient privacy computing fusion engine method and system, comprising: constructing medical data encryption and decryption system as foundation, standardization encapsulation is carried out to federated learning, homomorphic encryption and secure multi-party computation etc., and unified technical framework including data encryption interface, ciphertext calculation interface and result decryption interface is defined;Intelligent task analysis mechanism is used, and high-level medical analysis demand can be automatically decomposed into privacy computing subtask sequence, and multi-technology cooperation is realized through execution engine;At the same time, resource monitoring and performance analysis function is realized, running time technology switching and parameter tuning are supported, and self-optimizing execution environment is formed.Finally, through friendly application interface and visual display.The application solves the problems of single technical route, low computing efficiency, poor resource utilization and serious memory limitation in medical data privacy protection.

Description

Technical Field

[0001] This invention relates to the field of medical big data processing and privacy protection, specifically to a method and system for an efficient privacy computing fusion engine for medical data, which is mainly applied to scenarios of secure sharing and analysis of medical data across institutions. Background Technology

[0002] With the widespread application of big data in healthcare, the privacy protection and efficient utilization of medical data have become significant technological challenges. Medical data is characterized by high sensitivity, high dimensionality and heterogeneity, and cross-institutional dispersion, making it difficult to directly apply traditional data analysis methods to cross-institutional sharing and analysis scenarios.

[0003] Currently, privacy-preserving computation technology has been applied to some extent in the field of medical data, mainly including technical approaches such as federated learning, homomorphic encryption, and secure multi-party computation. Federated learning allows institutions to collaboratively train models without sharing the original data; homomorphic encryption supports direct computation on encrypted data; and secure multi-party computation allows multiple parties to jointly compute specific functions while protecting their own data privacy.

[0004] Existing technologies typically employ a single privacy-preserving computation approach to achieve secure sharing and analysis of medical data, such as medical data analysis systems based on homomorphic encryption or cross-institutional collaborative model training frameworks based on federated learning. These systems usually use fixed encryption schemes and computation protocols during implementation, resulting in relatively rigid system architectures and tight coupling between computational models and data processing flows.

[0005] However, these existing technologies have significant shortcomings: First, a single technical approach is difficult to adapt to the complex scenarios and diverse computing needs of medical data; second, existing encryption algorithms are inefficient and computationally expensive when processing high-dimensional medical data; third, there is a lack of adaptive technology selection mechanisms for different types of medical data and computing tasks, resulting in poor system performance and resource utilization; and finally, existing systems face serious memory limitations and computational efficiency problems when processing large-scale, high-concurrency medical data modeling. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for an efficient privacy-preserving computing fusion engine for medical data, aiming to solve the problems of single technical approach, low computing efficiency, poor resource utilization, and severe memory limitations in medical data privacy protection.

[0007] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0008] This invention provides a method for an efficient privacy-preserving computation fusion engine for medical data, comprising:

[0009] Based on the characteristics of multi-source medical data, a mapping relationship between state identifiers and prefix tree nodes is constructed through a prefix tree data structure to realize a read-write lock mechanism that supports multi-threaded concurrent access and a dynamic memory allocation strategy, thus obtaining the state management module.

[0010] Based on the state management module, the performance characteristics of various privacy computing technologies under different medical data scenarios are analyzed. The computing tasks are classified and modeled as optimization problems according to data scale, computing complexity and privacy requirements. An approximate algorithm is designed to integrate data types, computing tasks and system resource states to generate an adaptive privacy computing technology combination strategy.

[0011] Based on the adaptive privacy computing technology combination strategy, the privacy computing protocol is decomposed into basic operational elements and a mathematical model is constructed. The multidimensional weights of the computational cost, communication overhead and security strength of the protocol elements are defined. A solution algorithm is designed to select the minimum cost combination of protocol elements that meets the preset security threshold, and the optimized privacy computing protocol framework is obtained.

[0012] Based on the optimized privacy computing protocol framework, privacy sensitivity analysis and classification of medical data are performed, a hybrid encryption strategy is designed, and a distributed key generation, distribution and management system is implemented to obtain a medical data encryption and decryption system.

[0013] Based on the aforementioned medical data encryption and decryption system, various privacy computing technologies are standardized and encapsulated, and a unified interface is defined. An execution engine is built to achieve collaborative scheduling and execution of multiple technologies, resulting in a medical data privacy computing fusion system.

[0014] Furthermore, the mapping relationship between state identifiers and prefix tree nodes is constructed using a prefix tree data structure to implement a read-write lock mechanism that supports multi-threaded concurrent access and a dynamic memory allocation strategy, resulting in a state management module, including:

[0015] The data types, data scales, and access patterns of the multi-source medical data are analyzed. The state space generated during the privacy computation process is analyzed to determine the state characteristics and distribution patterns. A prefix tree structure suitable for privacy computation of medical data is designed, including node representation, path encoding, and tree balancing strategies, resulting in a prefix tree structure design scheme.

[0016] Based on the aforementioned prefix tree structure design scheme, a mapping relationship between state identifiers and prefix tree nodes is constructed to achieve efficient storage and retrieval of state information, resulting in a state-prefix tree mapping system.

[0017] Based on the state-prefix tree mapping system, for scenarios where multiple medical institutions simultaneously initiate federated learning training tasks and homomorphic encryption computation tasks, a read-write lock-based concurrency control mechanism is implemented to support multi-threaded simultaneous reading and mutually exclusive writing to prefix tree nodes. A dynamic memory allocation strategy is implemented to adaptively adjust the memory pool capacity according to the size of the state space, resulting in a concurrency-optimized prefix tree state management module.

[0018] Based on the concurrent optimization prefix tree state management module, the path compression technology of the prefix tree is implemented to merge paths with a single child node, and the incremental persistence mechanism is implemented to write only the changed prefix tree nodes to the storage medium, thereby reducing storage overhead and supporting breakpoint recovery of large-scale medical models, thus obtaining the state management module.

[0019] Furthermore, the analysis of the performance characteristics of various privacy-preserving computing technologies under different medical data scenarios classifies computing tasks according to data scale, computational complexity, and privacy requirements, and models them as optimization problems. An approximate algorithm is designed to integrate data types, computing tasks, and system resource states to generate an adaptive privacy-preserving computing technology combination strategy, including:

[0020] Based on the state management module, the state data of each privacy computing technology during the historical execution process is obtained, and the execution time, memory usage and computational accuracy characteristics of federated learning, homomorphic encryption and secure multi-party computation under different medical data types and computing scenarios are analyzed to obtain a technology characteristic evaluation matrix.

[0021] Based on the aforementioned technical characteristic evaluation matrix, medical data computing tasks are classified according to data scale, computational complexity, and privacy requirements, establishing a task classification system and obtaining a medical computing task category model.

[0022] Based on the aforementioned medical computing task category model, the privacy computing technology selection problem is modeled as a non-uniform k-center problem, and a cost function and resource constraints are defined to obtain a mathematical optimization model.

[0023] Based on the mathematical optimization model, an approximate algorithm for minimizing resources is designed and implemented to solve the non-uniform k-center problem, resulting in a technology selection algorithm.

[0024] Based on the aforementioned technology selection algorithm, taking into account data type, computational task, and system resource status, the optimal privacy computing technology combination strategy is dynamically generated, resulting in the adaptive privacy computing technology combination strategy.

[0025] Furthermore, the privacy computation protocol is decomposed into basic operational elements and a mathematical model is constructed. Multidimensional weights for the computational cost, communication overhead, and security strength of each protocol element are defined. A solution algorithm is designed to select the minimum cost combination of protocol elements that satisfies a preset security threshold, resulting in an optimized privacy computation protocol framework, including:

[0026] Based on the federated learning, homomorphic encryption and secure multi-party computation technologies determined in the adaptive privacy computing technology combination strategy, the privacy computing protocols corresponding to each technology are decomposed into basic operation elements such as key exchange, data encryption, ciphertext computation and result decryption, and a protocol element library is established to obtain a protocol element set.

[0027] Based on the set of protocol elements, an uncertain matroid model is constructed. The protocol elements are used as elements in the matroid, and the differential privacy protection strength to prevent data leakage and the computational efficiency requirements of protocol execution are used as independence constraints of the matroid, thus obtaining the matroid mathematical model.

[0028] Based on the matroid mathematical model, the computational cost (including the number of CPU cycles), communication overhead (including the amount of data transmitted), and security strength (including the differential privacy parameter ε) are defined for each protocol element. A comprehensive evaluation system is established to obtain the element weight matrix.

[0029] Based on the element weight matrix, a minimum weight basis solution algorithm is designed and implemented. The algorithm selects the protocol element combination with the differential privacy parameter ε value less than the preset security threshold and the minimum total computation cost and communication overhead from the protocol elements to obtain the protocol minimum weight basis.

[0030] Based on the minimum weight basis of the protocol, a protocol optimization framework is constructed to realize the dynamic combination and formal verification of protocol elements, thereby obtaining the optimized privacy computing protocol framework.

[0031] Furthermore, a privacy sensitivity analysis and classification of medical data are performed, a hybrid encryption strategy is designed, and a distributed key generation, distribution, and management system is implemented to obtain a medical data encryption and decryption system, including:

[0032] Based on the optimized privacy computing protocol framework, a privacy sensitivity quantification analysis is performed on patient identity information, diagnostic records and genetic data in medical data. The data is divided into three levels of sensitivity from high to low: high sensitivity, medium sensitivity and low sensitivity. A multi-level protection strategy is established to obtain a data sensitivity classification model.

[0033] Based on the data sensitivity classification model, a hybrid encryption strategy combining symmetric and asymmetric encryption is designed. Different encryption strengths are selected for data with different sensitivities to obtain the hybrid encryption strategy.

[0034] Based on the hybrid encryption strategy and the computational cost constraints in the optimized privacy computing protocol framework, the key expansion rounds and block mode parameters of the encryption algorithm are optimized to obtain the optimized encryption algorithm set.

[0035] Based on the optimized set of encryption algorithms, a distributed key generation, distribution and management system is designed and implemented to ensure key security in a multi-organization collaborative environment, thus obtaining a key management system.

[0036] Based on the key management system, the encryption and decryption functions are encapsulated into a unified interface and service, supporting calls from various privacy computing modules, thus obtaining the medical data encryption and decryption system.

[0037] Furthermore, the system standardizes and encapsulates various privacy-preserving computation technologies and defines a unified interface, constructs an execution engine to achieve collaborative scheduling and execution of multiple technologies, resulting in a medical data privacy-preserving computation fusion system, including:

[0038] Based on the API interface provided by the medical data encryption and decryption system, federated learning, homomorphic encryption, and secure multi-party computation are standardized and encapsulated. The data input and output of each technology are connected to the encryption and decryption service of the medical data encryption and decryption system. A unified interface is defined, including a data encryption interface, a ciphertext computation interface, and a result decryption interface, to obtain a standardized privacy computation module set.

[0039] Based on the standardized privacy computing module set, a parsing system for medical data analysis and model training tasks is implemented. The system receives user-submitted requests for disease prediction model training, multi-center clinical data statistical analysis, and genomic data association analysis. It calls the medical data encryption and decryption system to encrypt the original medical data and decomposes the requests for disease prediction model training, multi-center clinical data statistical analysis, and genomic data association analysis into a sequence of privacy computing sub-tasks to obtain a task execution plan.

[0040] Based on the task execution plan, a privacy computing execution engine core is constructed, integrating the key management function of the medical data encryption and decryption system to achieve multi-technology collaborative scheduling and execution, supporting data operations in encrypted state, thus obtaining the execution engine core system;

[0041] Based on the core system of the execution engine, a computing resource monitoring and performance analysis system is implemented to monitor the encryption and decryption overhead and key management status of the medical data encryption and decryption system, support runtime technology switching and parameter tuning, and obtain a self-optimizing execution environment.

[0042] Based on the self-optimizing execution environment, an application programming interface and a result visualization system are designed and implemented. The medical data encryption and decryption system is called to decrypt the calculation results and then display them, thus obtaining the medical data privacy calculation fusion system.

[0043] Furthermore, the state space generated during the privacy computation process is analyzed to determine state characteristics and distribution patterns. A prefix tree structure suitable for privacy computation of medical data is designed, including node representation, path encoding, and tree balancing strategies, resulting in a prefix tree structure design scheme, including:

[0044] Based on the data type, data scale and access pattern of the multi-source medical data, the state characteristics and distribution patterns in the state space of medical data are analyzed, the number of states, the frequency of state transitions and the state dependencies are determined, and a medical data state space characteristic model is obtained.

[0045] Based on the state space feature model of the medical data, a prefix tree node structure is designed, including a node data field, a child node pointer array, and a compression tag. A variable-length encoding scheme is designed for the path encoding, and an AVL tree balancing strategy is designed to maintain the query efficiency of the prefix tree, thus obtaining the prefix tree structure design scheme.

[0046] Furthermore, based on the medical computing task category model, the privacy computing technology selection problem is modeled as a non-uniform k-center problem, and a cost function and resource constraints are defined to obtain a mathematical optimization model, including:

[0047] Based on the medical computing task category model, all classified medical data computing tasks are extracted to form a medical data computing task set. Each task in the medical data computing task set is defined as a feature vector containing data scale, computational complexity, and privacy requirements. Privacy computing technologies, including federated learning, homomorphic encryption, and secure multi-party computation, are defined as a technology set, resulting in a task-technology mapping framework.

[0048] Based on the task-technology mapping framework, a distance function is defined to represent the cost of executing a task in the medical data computing task set using a technology in the technology set. A capacity constraint is defined for each technology in the technology set to represent the maximum number of tasks that each technology can process simultaneously, and a radius constraint is defined to represent the maximum acceptable execution cost of each technology, thus obtaining a set of constraints.

[0049] Based on the set of constraints, the objective function is constructed to minimize the largest task-technology distance among all tasks. The constraints are set as follows: the number of tasks assigned to each technology does not exceed the corresponding capacity constraint, and the distance between each task and the corresponding assigned technology does not exceed the radius constraint of the corresponding technology, thus obtaining the mathematical optimization model.

[0050] Further, based on the element weight matrix, the algorithm for designing and implementing the minimum weight basis solution selects a combination of protocol elements from the protocol elements whose differential privacy parameter ε is less than a preset security threshold and whose total computational cost and communication overhead are minimized, thus obtaining the protocol minimum weight basis, including:

[0051] Based on the computational cost, communication overhead, and security strength weight of each protocol element defined in the element weight matrix, a protocol element selection set is created by extracting all protocol elements in the element weight matrix and sorting all protocol elements in descending order of security strength to obtain a candidate element sequence.

[0052] Based on the candidate element sequence, a greedy strategy is used to sequentially select protocol elements that satisfy the differential privacy parameter ε value being less than a preset security threshold. The total computational cost and communication overhead of the currently selected element combination are calculated to obtain the initial protocol element combination.

[0053] Based on the initial combination of protocol elements, element replacement and combination optimization are performed iteratively. Under the premise of ensuring that the differential privacy parameter ε meets the security threshold, the total computational cost and communication overhead are continuously reduced. When the total computational cost and communication overhead cannot be reduced in multiple consecutive iterations, the iteration stops, and a local optimal combination of elements is obtained.

[0054] Based on the locally optimal combination of elements, the independence constraints and matroid properties of the selected protocol elements are verified to ensure that the combination of elements meets the requirements of protocol integrity and security, and the minimum weight basis of the protocol is obtained.

[0055] This invention also provides a high-efficiency privacy-preserving computing fusion engine system for medical data, comprising:

[0056] The state management module is used to construct a mapping relationship between state identifiers and prefix tree nodes based on the characteristics of multi-source medical data through a prefix tree data structure, and to implement a read-write lock mechanism that supports multi-threaded concurrent access and a dynamic memory allocation strategy, thus obtaining the state management module.

[0057] The adaptive technology selection module is used to analyze the performance characteristics of various privacy computing technologies in different medical data scenarios based on the state management module, classify computing tasks according to data scale, computing complexity and privacy requirements and model them as optimization problems, design approximate algorithms to integrate data types, computing tasks and system resource states, and generate an adaptive privacy computing technology combination strategy.

[0058] The protocol optimization module is used to decompose the privacy computing protocol into basic operational elements and construct a mathematical model according to the adaptive privacy computing technology combination strategy, define the multi-dimensional weights of the computation cost, communication overhead and security strength of the protocol elements, design a solution algorithm to select the minimum cost protocol element combination that satisfies the preset security threshold, and obtain the optimized privacy computing protocol framework.

[0059] The encryption / decryption module is used to perform privacy sensitivity analysis and classification on medical data based on the optimized privacy computing protocol framework, design a hybrid encryption strategy, and implement a distributed key generation, distribution, and management system to obtain a medical data encryption / decryption system.

[0060] A privacy computing execution engine is used to standardize and encapsulate various privacy computing technologies and define a unified interface based on the medical data encryption and decryption system. The execution engine is then used to achieve collaborative scheduling and execution of multiple technologies, resulting in a medical data privacy computing fusion system.

[0061] The advantages and positive effects of this invention are:

[0062] 1. By using a prefix tree instead of a hash map to store the state space for medical privacy computations, the memory limitation problem in medical big data modeling is solved. This method is particularly suitable for complex medical models such as Clinical Relationship Networks (CRNs) and reduces memory usage.

[0063] 2. The selection of medical privacy computing technology is modeled as a non-uniform k-center problem. An adaptive technology selection mechanism is implemented through an approximate algorithm that minimizes resources, enabling the system to automatically select the optimal privacy computing scheme based on the type of medical data, computing tasks, and performance requirements, thereby improving computing efficiency.

[0064] 3. An optimization framework for privacy computing protocols is constructed using the theory of uncertain matroids. The minimum weight basis verification mechanism is used to screen the minimum cost protocol combination that meets security requirements, thereby reducing computation and communication overhead and lowering system resource consumption while ensuring security.

[0065] 4. Design a hybrid encryption mechanism tailored to the characteristics of medical data, combining symmetric and asymmetric encryption technologies to achieve graded protection of data sensitivity, improving encryption and decryption efficiency while ensuring high security, and supporting medical data analysis in encrypted form;

[0066] 5. Construct a privacy-preserving computation execution engine that integrates multiple technologies to achieve unified scheduling and collaborative work of technologies such as federated learning, homomorphic encryption, and secure multi-party computation, supporting the secure sharing and analysis of cross-institutional medical data that is "usable but not visible". Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is a flowchart illustrating the efficient privacy-preserving computing fusion engine method for medical data according to the present invention.

[0069] Figure 2 This is a flowchart illustrating the workflow of the adaptive technology selection module in this invention.

[0070] Figure 3 This is a schematic diagram of the structure of the efficient privacy computing fusion engine system for medical data according to the present invention. Detailed Implementation

[0071] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0072] like Figure 1 As shown, this invention provides a method for an efficient privacy-preserving computation fusion engine for medical data, comprising:

[0073] Step S1: Based on the characteristics of multi-source medical data, a mapping relationship between state identifiers and prefix tree nodes is constructed through a prefix tree data structure to implement a read-write lock mechanism that supports multi-threaded concurrent access and a dynamic memory allocation strategy, thus obtaining the state management module.

[0074] Step S2: Based on the state management module, analyze the performance characteristics of various privacy computing technologies in different medical data scenarios, classify computing tasks according to data scale, computing complexity and privacy requirements and model them as optimization problems, design approximate algorithms to integrate data types, computing tasks and system resource states, and generate an adaptive privacy computing technology combination strategy.

[0075] Step S3: Based on the adaptive privacy computing technology combination strategy, decompose the privacy computing protocol into basic operation elements and construct a mathematical model. Define the multidimensional weights of the computation cost, communication overhead and security strength of the protocol elements. Design a solution algorithm to select the minimum cost protocol element combination that satisfies the preset security threshold, and obtain the optimized privacy computing protocol framework.

[0076] Step S4: Based on the optimized privacy computing protocol framework, perform privacy sensitivity analysis and classification on medical data, design a hybrid encryption strategy and implement a distributed key generation, distribution and management system to obtain a medical data encryption and decryption system;

[0077] Step S5: Based on the medical data encryption and decryption system, standardize and encapsulate various privacy computing technologies and define a unified interface. Construct an execution engine to achieve multi-technology collaborative scheduling and execution, thereby obtaining a medical data privacy computing fusion system.

[0078] In one embodiment of the present invention, step S1 further includes:

[0079] Step S11: Analyze the data types, data scales, and access patterns in the multi-source medical data features, analyze the state space generated during the privacy computation process, determine the state characteristics and distribution patterns, design a prefix tree structure suitable for privacy computation of medical data, including node representation, path encoding, and tree balancing strategies, and obtain the prefix tree structure design scheme.

[0080] First, we conduct an in-depth analysis of the characteristics of multi-source medical data. Multi-source medical data typically includes various types such as structured electronic health records (EHRs), semi-structured medical imaging data, and unstructured medical text records. In terms of data scale, clinical data is usually high-dimensional, with a single patient record potentially containing thousands of features, while genomic data has even higher dimensionality, with a single sample potentially containing millions of feature points. Regarding access patterns, the frequency of reading medical data is generally higher than the frequency of writing it. For example, in the training of disease prediction models, the frequency of model parameter updates is far lower than the frequency of data reading, exhibiting a typical read-heavy, write-light characteristic.

[0081] This step establishes a systematic analytical framework for the state space generated during privacy-preserving computation. In medical data privacy computation, the state space refers to the set of all intermediate states that need to be maintained during algorithm execution, including model parameters, gradient information, and encrypted intermediate values. Through analysis of typical privacy-preserving computation scenarios such as federated learning, homomorphic encryption, and secure multi-party computation, three key characteristics of the state space in medical data privacy computation are identified: state space sparsity, access locality, and hierarchical structure. State space sparsity means that the number of states actually used is very small, usually less than 1%, in the theoretically possible set of states; access locality is manifested in the fact that related states are usually close to each other in the identifier space; and hierarchical structure is reflected in the clear hierarchical dependencies between states, such as the association between model parameters and their corresponding gradients.

[0082] Based on the above analysis, characteristic models of the number of states, state transition frequency, and state dependencies were determined. In medical data privacy computation, the number of states is directly proportional to the data dimension and the number of participants; the state transition frequency exhibits periodic changes during model training and is related to the number of iterations; state dependencies can be formalized as a directed acyclic graph (DAG) structure, describing the evolutionary paths between states. These characteristics collectively constitute the characteristic model of the medical data state space.

[0083] Then, based on the medical data state space feature model, a prefix tree node structure is designed, including a node data field, a child node pointer array, and a compression tag. The prefix tree node structure consists of three parts: (1) a data field, which is used to store the state information associated with the node and supports variable-length state data; (2) a child node pointer array, which is implemented using a dynamic array and automatically adjusts its size according to the branch factor to avoid memory waste; and (3) a compression tag, which is used to identify whether the node participates in path compression and the length information of the compressed path.

[0084] In the path coding scheme design, variable-length coding technology is adopted, adaptively selecting the coding length based on the characteristics of the state identifier. Variable-length coding is a technique that assigns different length codes based on the probability of data occurrence. In this scheme, high-frequency access states are assigned shorter codes, and low-frequency access states are assigned longer codes, which matches the uneven distribution characteristics of medical data state access. The coding scheme uses a prefix rule to ensure the uniqueness of decoding, that is, no code is a prefix of another code.

[0085] To ensure the query efficiency of the prefix tree in dynamic environments, this scheme employs an AVL tree balancing strategy. An AVL tree is a self-balancing binary search tree where the height difference between the left and right subtrees of any node does not exceed 1, ensuring that the time complexity of the search operation is O(log n) in the worst case. In this prefix tree structure, the set of child nodes of each node forms an AVL tree. Left rotation, right rotation, and double rotation operations are used to maintain the tree's balance, ensuring that even with a large number of state insertions and deletions, the query efficiency remains at the logarithmic level.

[0086] Through the above design, a prefix tree structure design scheme specifically suitable for medical data privacy computing was finally formed. This scheme can efficiently handle the sparse, highly localized, and hierarchical state space of medical data.

[0087] Step S12: Based on the aforementioned prefix tree structure design scheme, construct the mapping relationship between state identifiers and prefix tree nodes to achieve efficient storage and retrieval of state information, thereby obtaining a state-prefix tree mapping system.

[0088] Based on the prefix tree structure designed in the aforementioned steps, a mapping system from state identifiers to prefix tree nodes is implemented. First, the definition of a state identifier is clarified: In medical data privacy computing, a state identifier is a string or numerical value that uniquely identifies a specific state during the computing process. It is usually composed of multiple fields, such as "participant ID-data type-feature index-timestamp", etc.

[0089] The core of the mapping system is to convert state identifiers into paths in a prefix tree and store state information at the end nodes of the corresponding paths. The specific implementation process includes three key steps: path encoding, node localization, and state storage.

[0090] The path encoding stage employs a recursive segmented encoding algorithm, decomposing the state identifier into multiple parts, each corresponding to a level of the prefix tree. The algorithm first analyzes the semantic structure of the identifier, identifying its natural segmentation points, such as field separators or hierarchical boundaries. Then, it applies variable-length encoding rules to each segment to generate a path encoding sequence. For example, for the identifier "hospital_A:patient_123:feature_456:20220315", possible segments are ["hospital_A", "patient_123", "feature_456", "20220315"]. Each segment, after encoding, forms the complete path.

[0091] The node localization phase implements an efficient pathfinding mechanism. Given a path encoding sequence, child nodes are searched layer by layer, starting from the root node of the prefix tree. To improve search efficiency, a two-level index structure is implemented on the child node set of each node: the first level is a hash index based on the first character of the child node, used to quickly narrow down the search range; the second level is an AVL tree structure, used to accurately locate child nodes with the same first character. This two-level index structure optimizes the average time complexity of node search to between O(1) and O(log m), where m is the number of sibling child nodes.

[0092] The state storage phase is responsible for storing or updating state information at the located nodes. An incremental storage strategy is employed, saving only the changed portions of the state, rather than the complete state data. For time-series medical data states, such as parameter changes during model iteration, a differential encoding mechanism is implemented, storing only the differences between adjacent states, significantly reducing storage space requirements. Furthermore, pluggable data compression modules are provided for different types of state data, supporting both lossless and lossy compression methods, allowing users to choose according to accuracy requirements and storage constraints.

[0093] By encoding state identifiers along paths and storing and retrieving them along prefix tree paths, efficient management of state information is achieved. Compared to traditional hash table mapping, this scheme has three significant advantages: First, it reduces memory fragmentation. Hash tables typically generate a large amount of memory fragmentation during dynamic expansion, while the prefix tree structure can reduce this fragmentation through path sharing and compression. Second, it reduces redundant data. State identifiers with similar prefixes share storage paths in the prefix tree, reducing duplicate information. Third, it increases the storage density of state information. Tests show that in typical medical data privacy computing scenarios, the prefix tree structure can reduce memory usage by approximately 40% compared to hash table storage.

[0094] Through the above implementation, a complete state-prefix tree mapping system was finally formed, providing basic support for subsequent concurrency control and memory management optimization.

[0095] Step S13: Based on the state-prefix tree mapping system, for scenarios where multiple medical institutions simultaneously initiate federated learning training tasks and homomorphic encryption computation tasks, implement a read-write lock-based concurrency control mechanism to support multi-threaded simultaneous reading and mutually exclusive writing to prefix tree nodes, and implement a dynamic memory allocation strategy to adaptively adjust the memory pool capacity according to the size of the state space, thereby obtaining a concurrency-optimized prefix tree state management module.

[0096] Based on the state-prefix tree mapping system constructed using the aforementioned steps, concurrency control and memory management optimizations are implemented for high-concurrency medical data processing scenarios. Multi-institutional concurrency is a common scenario in medical data privacy computing, such as multiple hospitals simultaneously participating in federated learning model training, or multiple research institutions conducting collaborative data analysis using homomorphic encryption. These scenarios place high demands on the concurrent processing capabilities of the state management module.

[0097] First, regarding concurrent access control for the prefix tree, this step implements a multi-granularity read-write lock mechanism. A read-write lock is a concurrency control mechanism that allows multiple readers to access the tree simultaneously while requiring exclusive access for writers, making it suitable for the read-heavy, write-light characteristics of medical data privacy computing. The innovation of this solution lies in its multi-granularity lock control, refining the lock scope from the entire tree to different levels and branches: root node locks control access to the entire tree; level locks control access to specific levels; branch locks control access to a subtree; and node locks control access to individual nodes. This multi-granularity lock design significantly improves concurrency because most operations only require locking a portion of the tree.

[0098] Lock acquisition follows a top-down order, meaning the upper-level lock is acquired first, followed by the lower-level lock, which avoids deadlock problems. Furthermore, lock escalation and degrade operations support dynamic adjustment of lock granularity. For example, when an operation expands from reading a few nodes to reading the entire tree, multiple node locks can be upgraded to a single root node read lock, improving efficiency and reducing lock overhead.

[0099] To further improve concurrency performance, this solution implements an optimistic concurrency control mechanism as a supplement to read-write locks. Optimistic concurrency control assumes that conflicts rarely occur, allowing transactions to execute first and verifying for conflicts before committing. In prefix tree operations, the version number of each node is recorded. During read operations, the version is recorded, and before writing, it is checked whether the version has changed. If it hasn't changed, writing is allowed and the version number is updated; otherwise, a retry or error is reported. This mechanism is particularly suitable for short-term read operations in medical data, such as parameter queries during model inference.

[0100] Secondly, regarding memory management, this step implements a dynamic memory allocation strategy, adaptively adjusting the memory pool capacity based on the state space size. Dynamic memory allocation is a technique that automatically adjusts memory allocation according to runtime needs, which is particularly important in healthcare privacy computing because the state space sizes of different computing tasks vary significantly, and static allocation can easily lead to memory waste or insufficiency.

[0101] This solution employs a layered memory pool design: a global memory pool handles the allocation and release of large memory blocks; local memory pools provide fast allocation for objects of specific sizes; and a temporary memory pool processes short-lived temporary objects. The memory allocation strategy combines pre-allocation and dynamic growth: a certain amount of memory is pre-allocated based on configuration at startup; during operation, memory usage is monitored, and memory expansion is triggered when it exceeds a preset threshold (e.g., 85%); and memory contraction is triggered when the usage remains below another threshold (e.g., 50%) for an extended period.

[0102] To optimize the memory layout of prefix tree nodes, this solution implements node pooling technology. Node pooling centrally manages nodes of similar size, reducing memory fragmentation and improving cache hit rate. Specifically, fixed-size node blocks are pre-allocated, each containing multiple prefix tree nodes; adjacent nodes are physically adjacent in memory, improving spatial locality; after a node is released, it is not immediately returned to the operating system but is marked as available for subsequent allocation and reuse. Tests show that node pooling technology can reduce memory allocation overhead by up to 30% in large-scale medical model training scenarios.

[0103] By combining multi-granularity read-write locks and a dynamic memory allocation strategy, this step implements a prefix tree state management module suitable for high-concurrency medical data processing scenarios. The read-write lock mechanism allows multiple read operations to occur simultaneously, while write operations require mutually exclusive access, adapting to the read-heavy and write-light characteristics of medical data privacy computing; the dynamic memory allocation strategy can automatically adjust the memory pool capacity according to the size of the state space of the actual computing task, avoiding memory waste or insufficient memory problems.

[0104] Step S14: Based on the concurrently optimized prefix tree state management module, implement the path compression technology of the prefix tree to merge paths with a single child node, implement the incremental persistence mechanism to write only the changed prefix tree nodes to the storage medium, reduce storage overhead and support breakpoint recovery of large-scale medical models, and obtain the state management module.

[0105] The concurrent optimization prefix tree state management module, implemented based on the aforementioned steps, further improves storage efficiency and system reliability through path compression and incremental persistence, thus completing the final state management module construction.

[0106] First, a path compression technique for prefix trees is implemented, merging paths with a single child node. Path compression is an optimization technique that reduces tree depth and the number of nodes by merging intermediate nodes. In this scheme, it is specifically designed to compress single-chain paths in prefix trees. A single-chain path refers to a series of nodes, where each node (except the last one) has only one child node. This type of path is common in the state space of medical data, such as the consecutive numerical portion of a patient ID sequence.

[0107] The path compression implementation adopts an adaptive triggering mechanism, and performs compression operations in the following three situations: (1) Periodic maintenance, scanning the entire tree at the configured time interval, identifying and compressing single-chain paths; (2) Threshold trigger, performing compression when the average depth of the tree or the total number of nodes exceeds the preset threshold; (3) Operation trigger, performing local compression when frequent insertions cause significant changes in the tree structure.

[0108] The core of the compression algorithm is a depth-first search (DFS) traversal of the tree structure to identify and merge singly linked paths. The merging process includes: creating new compressed nodes, concatenating and storing all characters along the path; maintaining the original split point information to support subsequent precise searches; and updating the parent node pointer to bypass the compressed node and directly point to the merged node. To ensure concurrency safety, the compression operation needs to acquire an exclusive lock on the corresponding branch and update the relevant indexes after compression is complete.

[0109] The advantages of path compression technology are multifaceted: first, it reduces the depth of the tree, lowering the average path length for query operations; second, it reduces the number of nodes, saving memory space; and third, it improves caching efficiency because more relevant data can be loaded into the cache line at once. Actual tests show that for a typical medical data state space, path compression can reduce the number of nodes by 20%-35% and improve query performance by 15%-25%.

[0110] Secondly, an incremental persistence mechanism is implemented, writing only the changed prefix tree nodes to the storage medium. Incremental persistence is a storage technology that saves only the changed parts rather than the entire set of data. In this solution, it is used to efficiently save the prefix tree state and support system recovery and breakpoint resume functionality.

[0111] The incremental persistence mechanism is designed based on a change tracking system, which maintains a modification log table that records the nodes affected by each operation and their modification type (add, update, delete). The persistence process consists of three layers: real-time log, which records the transaction log of all modification operations and supports crash recovery; incremental snapshot, which periodically saves snapshot data of changed nodes and supports incremental backup; and full snapshot, which saves the complete tree structure during major version updates or when the system is idle, serving as a basic recovery point.

[0112] To improve persistence efficiency, the system implements batch processing and asynchronous write mechanisms. Modification operations are first recorded in a change buffer in memory. When the buffer reaches a preset size or a preset time interval has elapsed, the system writes the changes in batches to the storage medium. The write process is executed asynchronously in an independent thread, without blocking the main computation flow, and data consistency is ensured through Write-Ahead Logging (WAL) technology.

[0113] Incremental persistence mechanisms bring key value to privacy-preserving computation of medical data: on the one hand, by saving only the changed parts, storage overhead is significantly reduced, especially for scenarios such as incremental training and model fine-tuning; on the other hand, it supports breakpoint recovery for large-scale medical models, allowing the training process to continue from the point of interruption without having to start from the beginning, which is particularly important for long-running medical model training tasks.

[0114] This step completed the construction of the state management module through the implementation of path compression technology and incremental persistence mechanism. Path compression technology reduces the depth and number of nodes of the tree by merging paths with a single child node, further reducing memory consumption; the incremental persistence mechanism supports breakpoint recovery during medical model training, improving the stability and reliability of the system. The final state management module can efficiently handle large-scale state information in medical data privacy computation, providing a solid foundation for subsequent technology selection and protocol optimization.

[0115] In one embodiment of the present invention, step S2 involves analyzing the performance characteristics of various privacy computing technologies under different medical data scenarios based on the state management module, classifying computing tasks according to data scale, computational complexity, and privacy requirements, and modeling them as optimization problems. An approximate algorithm is designed to integrate data types, computing tasks, and system resource states to generate an adaptive privacy computing technology combination strategy.

[0116] Specifically, in step S21, the state data of each privacy computing technology during its historical execution is obtained based on the state management module, and the execution time, memory usage, and computational accuracy characteristics of federated learning, homomorphic encryption, and secure multi-party computation under different medical data types and computational scenarios are analyzed to obtain a technology characteristic evaluation matrix.

[0117] This step begins by collecting historical execution data for various privacy-preserving computation technologies through the state management module. This data includes performance records of the technologies in different medical scenarios. The state management module is a key component of the entire system, responsible for tracking and storing information about the execution process of privacy-preserving computation tasks, including task parameter configuration, resource usage, execution time, and result accuracy, forming a complete execution status record.

[0118] For federated learning techniques, the performance of different model architectures (such as CNN, RNN, and decision trees) on horizontal and vertical segmentation data was analyzed. The analysis revealed that for horizontally segmented medical image data, the execution time of federated learning is approximately linearly related to the number of participants. However, when the number of participants exceeds 20, communication overhead accelerates the performance degradation. Regarding memory usage, model complexity is the main influencing factor; a typical medical image classification model requires 2-5GB of memory per participant. In terms of accuracy, compared to centralized learning, federated learning suffers less than 2% accuracy loss when the data distribution is balanced, but when the data is highly imbalanced (e.g., rare case distribution), the accuracy loss can reach over 15%.

[0119] For homomorphic encryption technologies, the performance of partially homomorphic encryption (such as Paillier) and fully homomorphic encryption (such as CKKS and BFV) in medical data computation was evaluated. Analysis shows that Paillier encryption, when handling simple aggregation operations (such as weighted averages), keeps the processing time for tens of millions of patient records to the minute level, with a relatively moderate memory footprint (<8GB). In contrast, fully homomorphic schemes such as CKKS, while supporting complex operations (such as logistic regression iterations), increase the processing time for the same scale of data to the hour level, and significantly increase memory requirements (>32GB). Regarding accuracy, partially homomorphic schemes can maintain accurate calculations, while fully homomorphic schemes typically introduce less than 0.1% computational error due to noise management.

[0120] For secure multi-party computation techniques, the characteristics of different protocols, including obfuscated circuits, secret sharing, and unintentional transmission, in medical data processing were analyzed. Results show that the secret sharing scheme has the highest communication efficiency in scenarios with three or more parties, and statistical analysis tasks processing millions of patient records can be completed in minutes. Obfuscated circuits are suitable for complex but short-path computational logic, such as risk score calculation for a single patient across multiple institutions, but their performance rapidly declines with increasing circuit complexity. Regarding memory usage, the unintentional transmission protocol has high bandwidth requirements but relatively stable memory consumption. In terms of accuracy, integer-based schemes require preprocessing scaling, which may introduce truncation errors, while floating-point protocols maintain high accuracy at the cost of significantly reduced computational efficiency.

[0121] By integrating the above analysis results, a technology characteristic evaluation matrix was constructed. This is a multi-dimensional data structure containing three main dimensions: technology type (federated learning, homomorphic encryption, secure multi-party computation and its variants), medical data type (structured clinical data, medical imaging, genomic data, etc.), and computational task type (statistical analysis, predictive modeling, similarity search, etc.). Each cell in the matrix records key performance indicators for a specific technology-data-task combination, including execution time distribution, memory consumption curve, communication overhead, accuracy performance, and scalability characteristics. This matrix provides a data foundation and empirical reference for subsequent task classification and technology selection.

[0122] By utilizing historical execution data stored in the state management module, the performance of different privacy computing technologies in processing various types of medical data was systematically evaluated, including key indicators such as execution efficiency, resource consumption, and computational accuracy. This resulted in a technology characteristic evaluation matrix, providing data support for subsequent technology selection.

[0123] Step S22: Based on the technical characteristic evaluation matrix, medical data computing tasks are classified according to data scale, computational complexity, and privacy requirements to establish a task classification system and obtain a medical computing task category model.

[0124] This step, based on the technical characteristic evaluation matrix constructed in the preceding steps, systematically classifies medical data computing tasks, forming a structured task category model. This classification process takes into account the specific characteristics of medical data and computing tasks, providing a category basis for subsequent technology selection.

[0125] First, tasks are categorized based on data size. This classification considers not only the volume of data but also factors such as dimensionality and sparsity. Small-scale data (S) refers to datasets with a total size less than 10GB and fewer than 100,000 records, such as single-center clinical trial data; medium-scale data (M) refers to datasets with a total size of 10-100GB or 100,000-1,000,000 records, such as multi-center specialty case databases; large-scale data (L) refers to datasets with a total size exceeding 100GB or more than 1,000,000 records, such as data from a national health information platform. High-dimensional data (such as whole-genome data or whole-brain imaging) may be classified as medium- or large-scale even with a relatively small number of records.

[0126] Secondly, tasks are categorized based on computational complexity. Here, computational complexity refers to a comprehensive consideration of the algorithm's theoretical complexity and its actual execution characteristics. Low-complexity tasks (L) mainly include simple statistics (such as mean and median calculations), basic queries, and linear model training, with time complexities typically O(n) or O(n log n). Medium-complexity tasks (M) include general machine learning model training (such as decision trees and support vector machines), cluster analysis, and moderately complex image processing, with time complexities typically O(n²) or O(n log² n). High-complexity tasks (H) include deep neural network training, complex graph analysis, and genome sequence alignment, with time complexities potentially reaching O(n³) or higher. Furthermore, parallelization potential is also considered; tasks with high parallelization potential may have lower actual execution complexity even with higher theoretical complexity.

[0127] Next, the tasks are categorized based on privacy requirements. Privacy requirement assessments are based on a comprehensive consideration of data sensitivity, regulatory requirements, and potential risks. Low privacy requirement tasks (L) handle de-identified or aggregated data, such as anonymized public health statistics; medium privacy requirement tasks (M) handle limited sensitive information, such as de-identified but potentially re-identifiable clinical data; high privacy requirement tasks (H) handle highly sensitive information, such as genetic information, mental health records, or detailed health histories that directly identify individuals. The system combines HIPAA, GDPR, and other regulatory requirements with the organization's internal privacy policies for rating, and considers the specific requirements of different jurisdictions.

[0128] By combining these three core dimensions, 27 theoretical task categories (3×3×3) were identified. However, some combinations are rare in real-world medical environments (such as the combination of low privacy requirements and high complexity). Therefore, cluster analysis was used to merge similar categories, ultimately identifying 12 typical medical computing task categories.

[0129] These 12 typical categories include: Large-scale - High Complexity - High Privacy (LHH) category, such as multicenter genome association analysis; Large-scale - Medium Complexity - High Privacy (LMH) category, such as population-wide health risk prediction models; Medium-scale - Medium Complexity - High Privacy (MMH) category, such as patient stratification analysis based on sensitive diagnostic data; and Small-scale - High Complexity - High Privacy (SHH) category, such as rare disease precision treatment models. Each category is accompanied by a description of typical application scenarios and key characteristic indicators, such as data distribution characteristics, computational patterns, and accuracy requirements.

[0130] Finally, a complete medical computing task category model was constructed. This model not only includes task classification labels but also typical feature vectors, performance requirements, and preliminary assessments of applicable technologies for each task category. This model supports rapid classification and preliminary technology matching for new tasks, providing a framework for subsequent precise optimization. Through the category model, the system can quickly identify the characteristics of new tasks and connect them to historical experience, improving the accuracy and efficiency of technology selection.

[0131] Medical data computing tasks are classified according to dimensions such as data scale (small, medium, large), computational complexity (low, medium, high), and privacy requirements (low, medium, high). A task classification system is established, and typical task categories are identified, such as gene data analysis tasks with "large scale, high complexity, and high privacy" and clinical diagnosis prediction tasks with "medium scale, medium complexity, and high privacy".

[0132] Step S23: Based on the medical computing task category model, the privacy computing technology selection problem is modeled as a non-uniform k-center problem, and a cost function and resource constraints are defined to obtain a mathematical optimization model.

[0133] This step, based on the task category model established in the preceding steps, formalizes the privacy computing technology selection problem into a non-uniform k-center problem, a variant of the classic facility location problem. The difference between the non-uniform k-center problem and the standard k-center problem lies in the fact that each center (in this example, the privacy computing technology) has different capacity limitations and service radii, which better reflects the differences in actual technological characteristics.

[0134] First, a task-technology mapping framework was constructed, abstracting medical data computation tasks and privacy-preserving computation technologies as points and centers in a feature space, respectively. Each medical data computation task is represented as a feature vector t. i This includes key dimensions such as data scale, computational complexity, and privacy requirements. For example, a typical medical image classification task might be represented as t i =<0.7, 0.6, 0.8>, which correspond to medium to large-scale data, medium computational complexity, and high privacy requirements, respectively.

[0135] Similarly, each privacy computing technology is defined as an element h in the set of technologies. j This includes variations with different parameter configurations. For example, the federated learning technology set includes horizontal federated learning, vertical federated learning, and their different aggregation algorithms and differential privacy configurations; the homomorphic encryption technology set includes schemes such as Paillier and CKKS, and their different security parameter configurations; the secure multi-party computation technology set includes different implementations based on obfuscated circuits, secret sharing, etc. Each element in the technology set is also represented as a feature vector, describing its computational power, security features, and resource requirements.

[0136] Next, the distance function d(t) is defined. i , h j ), representing task t i Using technology h j The cost incurred during execution. This distance function is a weighted combination of multi-dimensional costs, including: computation time cost, reflecting the processing time required for task execution; memory consumption cost, representing memory usage during execution; communication overhead cost, quantifying data transmission requirements; accuracy loss cost, measuring deviation from the ideal result; and security risk cost, assessing potential privacy leakage risks. The weights of each cost dimension are dynamically adjusted according to the application scenario and user priorities. For example, the weight of computation time will increase in emergency room analysis scenarios, while the weight of security will increase in sensitive gene data analysis.

[0137] For each technology h j Two types of key constraints are defined: capacity constraint cap(h) j The radius constraint rad(h) represents the maximum number of tasks that the technology can process simultaneously, reflecting the limitations of computing resources and concurrent processing capabilities; j This represents the maximum acceptable task-technology distance, defining the boundary of the technology's applicability. When the "distance" between the task and the technology exceeds this threshold, it means that the technology is no longer suitable for handling the task, possibly due to insufficient performance, excessive resource consumption, or inadequate security guarantees.

[0138] Based on the above definitions, a mathematical optimization model for the non-uniform k-center problem is constructed. The objective function is set as minimizing the maximum task-technology distance among all tasks (minimax optimization), i.e., min max{d(t i , h j ) | t i Assigned to h j This objective function focuses on optimizing the worst-performing task-technology match in the system, thereby improving the overall system's performance floor.

[0139] The constraints include: each task must be assigned to one and only one technology; the number of tasks assigned to each technology cannot exceed its capacity constraint; the distance between a task and the assigned technology must not exceed the radius constraint of that technology; and global constraints to ensure that the total system resource consumption is within the range of available resources.

[0140] This mathematical optimization model transforms the abstract technology selection problem into a structured combinatorial optimization problem, which can be solved using mature optimization algorithms. The model specifically considers the unique needs of medical scenarios, such as strict privacy requirements, complex computational patterns, and limited resource environments, ensuring that the generated technology selection solutions are both high-performance and meet the specific constraints of medical applications.

[0141] Based on the medical computing task category model, the problem of selecting privacy computing technology is modeled as a non-uniform k-center problem. A distance function that comprehensively considers computing cost, communication overhead and security is defined, as well as constraints that reflect the technology's processing power and applicability. A mathematical optimization model is constructed with the goal of minimizing the maximum task-technology distance.

[0142] Step S24: Based on the mathematical optimization model, design and implement an approximate algorithm for minimizing resources to solve the non-uniform k-center problem and obtain the technology selection algorithm.

[0143] This step designs an efficient approximation algorithm for the non-uniform k-center optimization model constructed in the preceding steps. Since the non-uniform k-center problem is NP-hard, finding the exact global optimum is usually computationally infeasible for real-world medical task sets. Therefore, this step develops a performance-guaranteed approximation algorithm.

[0144] First, the unique structure and properties of the problem are analyzed. The technology selection problem for healthcare privacy computing has several advantageous characteristics: the set of technologies is relatively limited (usually no more than 100 configurations); task-technology compatibility typically exhibits a sparse pattern, meaning not all tasks can be handled by all technologies; and there are clear hierarchical and substitution relationships among technologies, which can be used for pruning the search space.

[0145] Based on these characteristics, an approximation algorithm combining a greedy strategy and local search was designed. This algorithm consists of five core stages:

[0146] In the initial technology selection phase, a subset of technologies is preliminarily determined for representative tasks within the task set. The algorithm first sorts all tasks by security requirements from highest to lowest, then, starting with the task with the highest security requirements, selects the highest-capacity technology that satisfies its radius constraint for each task. This greedy selection strategy prioritizes the needs of difficult tasks while minimizing the number of required technologies. For task clusters with similar security requirements, the algorithm considers the similarity between tasks and attempts to select the same technology for similar task groups, reducing technology switching overhead.

[0147] The task allocation phase assigns tasks to the initially selected technologies. This phase uses a modified Hungarian algorithm to minimize the total task-technology distance while ensuring radius constraints. The algorithm first constructs a task-technology bipartite graph, with edge weights corresponding to the distance values, and then searches for the optimal match. For tasks that cannot be allocated due to capacity limitations, the algorithm dynamically increases the technology set, selecting new technologies that cover these tasks and increase the minimum total cost.

[0148] The local optimization phase iteratively improves the initial allocation scheme. The algorithm implements a technology replacement strategy, attempting to replace the current technology with a better-performing one; a task redistribution strategy, changing task allocation to balance the load; and boundary task optimization, focusing on "boundary tasks" close to the technology radius constraint and finding more suitable technologies for them. After each optimization operation, the algorithm updates the total cost and checks whether all constraints are still satisfied.

[0149] The capacity balancing phase focuses on balancing technology utilization to avoid some technologies being overloaded while others are idle. The algorithm calculates the utilization rate of each technology (number of assigned tasks / capacity limit), identifying high-load (>90% capacity) and low-load (<30% capacity) technologies. Then, task migration is implemented, transferring boundary tasks from high-load technologies to compatible low-load technologies, or adjusting technology parameters to increase the capacity of high-demand technologies. This balancing process not only improves system stability but also provides a buffer for dynamic tasks.

[0150] The performance prediction and validation phase uses historical data and performance models to verify the actual performance of the optimization scheme. The algorithm simulates the execution of the scheme on historical data from the state management module, predicting key performance indicators such as average execution time, peak resource usage, and service quality. If the predicted performance is unsatisfactory, the algorithm adjusts the optimization parameters (such as cost weights or radius constraints) and re-executes the aforementioned phase. This prediction-based validation reduces the performance deviation of the scheme after actual deployment.

[0151] The approximation algorithm has a time complexity of O(m²n log n), where m is the number of techniques and n is the number of tasks, significantly lower than the exponential complexity of the exact algorithm. Tests on real-world medical datasets demonstrate that the algorithm can generate high-quality solutions for scenarios involving hundreds of tasks within seconds, with a deviation from the theoretical lower bound typically less than 20%, meeting practical application requirements.

[0152] Compared to pure greedy algorithms or random search methods, this algorithm reduces the average task execution time by 35% and improves resource utilization by 40% while satisfying all constraints, especially when handling mixed task loads (including tasks of varying sizes and complexities). Furthermore, the local search and iterative optimization mechanisms in the algorithm design enable it to maintain stable performance even in dynamic environments where task characteristics change rapidly.

[0153] For the non-uniform k-center problem, an approximate algorithm based on a combination of greedy strategy and local search is designed. This algorithm efficiently solves the technology selection problem and obtains an approximately optimal technology-task allocation scheme through five stages: initial technology selection, task allocation, local optimization, capacity balancing, and performance prediction and verification.

[0154] Step S25: Based on the technology selection algorithm, taking into account data type, computing task and system resource status, dynamically generate the optimal privacy computing technology combination strategy to obtain the adaptive privacy computing technology combination strategy.

[0155] This step, based on the technology selection algorithm implemented in the preceding steps, constructs a complete adaptive technology combination strategy system. Unlike static solutions, this system can sense environmental changes and dynamically adjust its strategy, achieving true adaptive optimization.

[0156] First, a multi-dimensional environmental perception mechanism was designed to continuously monitor key factors influencing technology selection. This mechanism includes three core monitoring modules: a data characteristic monitoring module that tracks the dynamic characteristics of medical data, including changes in data size, dimensionality growth, and distribution shifts; a task load monitoring module that analyzes the queue status, priority changes, and arrival patterns of computational tasks; and a resource status monitoring module that collects real-time data on system resource usage, including processor load, memory usage, network bandwidth, and storage I / O. This monitoring data, after preprocessing and feature extraction, is converted into an environmental state vector, which serves as input for adaptive decision-making.

[0157] Secondly, a context-aware decision engine was implemented, dynamically adjusting the technology combination strategy based on the current environmental state. The decision engine is based on a reinforcement learning framework, modeling the technology selection process as a Markov decision process. The state space is defined by the environmental state vector, and the action space includes operations such as technology selection adjustment, parameter configuration update, and resource reallocation. The reward function comprehensively considers execution efficiency, resource utilization, and stability, encouraging the minimization of resource consumption while meeting performance requirements.

[0158] A three-tiered strategy generation mechanism is adopted, dividing responsibilities according to the decision-making timescale and scope: The long-term strategy layer (time span: hours to days) is responsible for the basic structure of the technology portfolio, determining which core technologies should be deployed and maintained, and formulating strategies based on task trend prediction and resource planning; the medium-term strategy layer (time span: minutes to hours) manages the configuration of technical parameters and resource allocation, adjusting the operating parameters and resource quotas of each technology according to the current task queue and system load; the short-term strategy layer (time span: seconds to minutes) handles real-time task scheduling and technology switching decisions, selecting the most suitable technology instance for each newly arriving task. The three-tiered strategy works collaboratively to ensure that the system maintains optimal performance across all timescales.

[0159] A scenario-based adaptive rule engine has been implemented to handle special scenarios and emergency situations. The rule engine contains a series of condition-action rules, such as "IF system CPU utilization > 90% THEN route new tasks to low computationally intensive technologies" or "IF high-priority emergency task detected THEN pause low-priority tasks and release resources." These rules are defined by domain experts and continuously refined through practical experience to ensure that the system maintains reasonable behavior even under extreme conditions.

[0160] To support a smooth technology transition, a task migration protocol and a state preservation mechanism were designed. When the decision engine determines that an ongoing task needs to be migrated from one technology to another, the migration protocol ensures the safe transfer of intermediate states and data, minimizing migration costs and preventing data loss or privacy breaches. For long-running tasks that cannot be interrupted, a checkpointing mechanism is implemented to periodically save the execution state to support possible future migrations.

[0161] Long-term optimization is achieved through self-assessment and continuous learning. Execution history is recorded in the state management module, and historical data is analyzed regularly to evaluate decision quality and adjust internal models. This closed-loop learning enables adaptation to long-term changes in healthcare data processing patterns, such as seasonal variations, evolution of medical practices, or hardware upgrades.

[0162] In one embodiment of the present invention, such as Figure 2 As shown, in step S3, based on the adaptive privacy computing technology combination strategy, the privacy computing protocol is decomposed into basic operational elements and a mathematical model is constructed. The multidimensional weights of the computational cost, communication overhead, and security strength of the protocol elements are defined. A solution algorithm is designed to select the minimum cost protocol element combination that satisfies the preset security threshold, thereby obtaining the optimized privacy computing protocol framework.

[0163] Specifically, in step S31, based on the federated learning, homomorphic encryption, and secure multi-party computation technologies determined in the adaptive privacy computing technology combination strategy, the privacy computing protocols corresponding to each technology are decomposed into basic operational elements such as key exchange, data encryption, ciphertext computation, and result decryption, and a protocol element library is established to obtain a protocol element set.

[0164] Based on the adaptive privacy computing technology combination strategy constructed in the aforementioned embodiments, the underlying protocols of different privacy computing technologies are systematically decomposed, basic operational elements are extracted, and a structured protocol element library is formed. Protocol decomposition is a method of breaking down complex protocols into basic functional units, which facilitates modular analysis and optimization.

[0165] First, the protocol of federated learning technology is decomposed. Federated learning protocols mainly include two types: horizontal federated learning and vertical federated learning, each with several variations. Taking horizontal federated learning as an example, it can be decomposed into the following key elements: secure aggregation elements, responsible for the secure aggregation of model parameters from multiple parties, including variations such as SecureSum, SecureAverage, and threshold secret sharing; gradient encryption elements, including variations such as gradient encryption based on homomorphic encryption and gradient perturbation based on differential privacy; model distribution elements, responsible for the secure transmission of the global model from the central server to the participants; and parameter verification elements, verifying the legitimacy of parameters uploaded by participants to prevent malicious participants from disrupting model training. Vertical federated learning additionally includes entity alignment elements (such as the PSI protocol) and segmented computation elements (handling computations in different feature spaces).

[0166] Secondly, the homomorphic encryption protocol is decomposed. Homomorphic encryption protocols are divided into two main categories based on the types of operations they support: partially homomorphic and fully homomorphic. The system extracts the following key elements from these: key generation elements, including key generation, distribution, and management, covering different implementations of schemes such as RSA, Paillier, BGV, and CKKS; data encryption elements, which convert plaintext data into ciphertext format; homomorphic operation elements, which perform specific calculations on the ciphertext, including homomorphic addition, homomorphic multiplication, and homomorphic comparison; noise management elements, which handle the noise growth problem in fully homomorphic encryption; and ciphertext decryption elements, which convert the calculation result from ciphertext back to plaintext.

[0167] Next, the protocol of secure multi-party computation (MPC) technology is decomposed. MPC encompasses various implementation methods, including obfuscated circuits, secret sharing, and unintentional transmission. Key extracted elements include: circuit construction elements, which convert computational tasks into Boolean or arithmetic circuits; secret sharing elements, which enable data fragmentation and reconstruction, including variations such as Shamir sharing and additive sharing; unintentional transmission elements, which enable secure information exchange between parties without revealing their choices; secure comparison elements, which implement conditional judgment operations such as size comparisons; and random number generation elements, which provide the secure randomness required by the protocol.

[0168] In addition, a variety of common elements across different technologies were identified, such as secure channel establishment elements (including TLS / SSL protocols), authentication elements (based on PKI or zero-knowledge proofs), and protocol synchronization elements (ensuring the synchronization of multi-party protocol execution).

[0169] Through this systematic decomposition, an element library containing approximately 120 basic protocol elements was ultimately established. Each element was standardized and defined, including a functional description, interface specification, security attributes, and resource requirements. Dependencies between elements were also explicitly marked, forming a directed graph structure that indicates which elements must be used together or mutually exclusive. For example, certain homomorphic encryption elements must be used in conjunction with their corresponding key generation and decryption elements; while some differential privacy elements are mutually exclusive with other privacy-preserving elements and cannot be applied to the same data simultaneously.

[0170] To support subsequent analysis, each protocol element was also labeled with additional attributes such as technology maturity (research prototype, commercially available, or widely deployed), implementation complexity (simple, medium, or complex), and hardware acceleration potential (CPU optimization, GPU acceleration, or dedicated hardware support). This information helps in considering implementation difficulty and performance optimization potential during actual deployment.

[0171] This protocol decomposition method breaks down complex privacy-preserving computation technologies into basic building blocks, facilitating subsequent refined analysis and optimization, thus improving the flexibility and efficiency of protocol design. The set of protocol elements, serving as the foundation for constructing optimized protocols, will be further analyzed and selected in subsequent steps.

[0172] Based on the technologies selected in the adaptive privacy computing technology combination strategy, the federated learning, homomorphic encryption and secure multi-party computation protocol are decomposed into about 120 basic operational elements, covering key exchange, data encryption, ciphertext computation and result decryption functions at each stage of protocol execution, forming a standardized protocol element library, providing a modular construction foundation for subsequent optimization.

[0173] Step S32: Based on the set of protocol elements, construct an uncertain matroid model, taking the protocol elements as elements in the matroid, and taking the differential privacy protection strength to prevent data leakage and the computational efficiency requirements of protocol execution as independence constraints of the matroid, to obtain the matroid mathematical model.

[0174] The protocol element set constructed in the preceding steps is mapped onto the theoretical framework of uncertain matroids, establishing a mathematical model to describe the relationships and constraints between the elements. Uncertain matroids are mathematical tools in combinatorics used to describe the independence structure of a set of elements; in this embodiment, they are used to formalize the combinatorial optimization problem of the privacy computing protocol.

[0175] Uncertainty matroid models are an extension of standard matroid theory, capable of handling situations where elements have uncertain relationships. In privacy computing protocol optimization, this uncertainty stems from the fact that the security and efficiency characteristics of element combinations are not simply additive, but involve complex interactions. Uncertainty matroids, by probabilistically representing independence relationships, more accurately describe this complexity.

[0176] First, the basic structure of matroids is defined. The set E of protocol elements constructed by S1 is taken as the universe set of the matroid, where each protocol element e∈E is an element in the matroid. A matroid M is defined as an ordered pair (E,I), where I is a family of independent subsets of E, satisfying the following three matroid axioms: nonemptiness (… ∈I), heritability (if X∈I and Y X, then Y∈I), and commutativity (if X, Y∈I and |X|<|Y|, then there exists an element e∈Y\X such that X∪{e}∈I).

[0177] Two types of key independence constraints are defined to determine whether a subset of elements constitutes an independent set:

[0178] Differential privacy protection strength constraints ensure that the protocol provides sufficient privacy protection. Differential privacy is a strict definition of privacy that ensures the presence or absence of any single data point has a limited impact on the computational results by adding calibrated noise during data processing. The differential privacy parameter ε is used as a key indicator to measure the strength of privacy protection; a smaller ε value indicates stronger privacy protection. For a subset S of elements... If the combination of elements can guarantee ε(S)≤εmax (the preset maximum allowed ε value), then S meets the independence requirement in the privacy protection dimension. The calculation of the effective ε value of different element combinations takes into account the combination effect. Some elements (such as noise-adding techniques) will directly contribute to the ε budget, while other elements (such as secure channels) may not directly affect the ε value but provide support for overall security.

[0179] Computational efficiency constraints ensure that the protocol can execute under reasonable resource limitations. A computational efficiency function η is defined, mapping the element subset S to the interval [0,1]. η(S) = 1 represents ideal efficiency, and η(S) = 0 represents completely unacceptable efficiency. η(S) considers multiple factors such as total execution time, memory consumption, and communication rounds. For the element subset S... E, if η(S)≥ηmin (the preset minimum efficiency threshold), then S satisfies the independence requirement in the efficiency dimension.

[0180] Combining these two constraints, a family of independent subsets I is defined: I = {S} E |ε(S)≤εmax and η(S)≥ηmin and S is functionally complete}. Functional completeness requires that the subset of elements can support all the basic functions required for the execution of the complete protocol, including key management, data encryption, computation processing, and result decryption.

[0181] An uncertainty factor is introduced to describe the uncertainty of the effect of element combination. In actual protocol deployment, the precise performance and security of elements may fluctuate due to factors such as data characteristics, network conditions, and system load. The uncertainty of differential privacy parameters and efficiency of a given subset S of elements is described by probability distribution functions p(ε|S) and p(η|S). Based on these distributions, the constraint satisfaction probability is defined as: P(S satisfies the constraint) = P(ε(S)≤εmax)·P(η(S)≥ηmin).

[0182] Based on the theory of uncertain matroids, the basis and independent sets of matroids are further defined. A basis is a maximal independent set, i.e., a set whose independence cannot be maintained by adding any more elements. In this model, the basis corresponds to the maximal combination of protocol elements that is functionally complete and satisfies privacy and efficiency constraints. Independent sets are subsets of the basis; they may not be functionally complete, but they satisfy the constraints within the scope of their included functionality.

[0183] A mathematical representation of matroids was constructed, including independence criteria, rank functions, and eigenvector representations of the basis. These mathematical tools support subsequent weight analysis and optimization algorithm design. By mapping the protocol optimization problem to the matroid theoretical framework, optimal protocol configurations can be found using mature combinatorial optimization methods.

[0184] By mapping protocol elements to an uncertain matroid model, a mathematical framework is established to describe the relationships between elements. The differential privacy parameter ε and computational efficiency η are defined as independence constraints, requiring that effective protocol combinations must simultaneously satisfy both the privacy protection strength threshold and computational efficiency requirements. The uncertainty of the effect of element combinations is also considered, providing a theoretical basis for subsequent optimization.

[0185] Step S33: Based on the matroid mathematical model, define the computational cost (including CPU cycles), communication overhead (including data transmission volume), and security strength (including differential privacy parameter ε value) for each protocol element, establish a comprehensive evaluation system, and obtain the element weight matrix.

[0186] Based on the matroid model constructed in the preceding steps, multi-dimensional weights are defined for each element in the protocol element set, constructing a comprehensive evaluation system and forming an element weight matrix. This weight matrix provides a quantitative description of the element performance and security characteristics, providing a decision-making basis for subsequent optimization algorithms.

[0187] First, a weighting evaluation method for three core dimensions was designed:

[0188] The computational cost dimension assesses the computational resources required for the execution of protocol elements. Key metrics include CPU cycles (reflecting processor load), memory usage (in MB), and computation latency (in milliseconds). A multi-level evaluation method is employed to obtain these metrics: for mature elements, precise values ​​are obtained through actual measurements on standard hardware platforms; for newly developed elements, estimations are made by combining theoretical complexity analysis and prototype testing; for elements dependent on data scale, a parameterized cost model is constructed to represent the functional relationship between cost and data volume. For example, the CPU cycles for some encryption elements can be expressed as O(n log n) complexity, where n is the number of data records. Specific cost values ​​are pre-calculated for common medical data scales (e.g., thousands, tens of thousands, millions of records).

[0189] The communication overhead dimension quantifies the network traffic generated by protocol elements during execution. Key metrics include data volume transmitted (in KB), number of communication rounds (affecting latency), and bandwidth requirements (in Mbps). The communication overhead assessment considers the characteristics of different network environments (e.g., intra-hospital LANs, inter-institutional WANs, and the public internet). For elements requiring multi-party interaction (e.g., secret sharing elements in secure multi-party computation), the impact of communication topology (e.g., star, ring, or fully connected) on overhead is also analyzed. In particular, the additional overhead of communication failure recovery mechanisms, which is crucial for the stability of medical applications, is evaluated.

[0190] The security strength dimension assesses the privacy protection capabilities provided by each element. The core indicator is the differential privacy parameter ε, which represents the strength of privacy protection (the smaller the ε, the stronger the protection). For elements that directly implement differential privacy (such as those with added Laplace noise or exponential noise), the ε value is directly determined by the algorithm parameters; for elements based on other security mechanisms (such as homomorphic encryption or secret sharing), they are mapped to the corresponding ε value through security equivalence analysis. In addition to the differential privacy parameter, auxiliary security indicators such as computational security (based on the strength of cryptographic assumptions), adversarial models (such as semi-honest or malicious adversaries), and side-channel resistance are also considered to form a comprehensive security score.

[0191] Based on these three core dimensions, an element weight matrix W is constructed. For each element e in the protocol element set E, the matrix W contains a row vector W(e) = [Wcomp(e), Wcomm(e), Wsec(e)], which represents the computation cost weight, communication overhead weight, and security strength weight, respectively (converted to cost representation, i.e., 1 / security strength, so that all dimensions are as small as possible).

[0192] To address the dimensional differences in indicators across different dimensions, standardization was implemented. Computational costs and communication overhead were mapped to the [0,1] interval using linear normalization; security strength was mapped to the [0,1] interval using an exponential transformation, ensuring that small differences in ε produce more significant weight differences in the high security requirement region (ε close to 0).

[0193] A context-weighting mechanism is also introduced to adjust the relative importance of each dimension according to the application scenario. For example, in emergency medical analysis scenarios, the weight of computational efficiency will be increased; while when processing highly sensitive genetic data, the weight of safety intensity will be increased. This adjustment is achieved through the weight vector α=[αcomp, αcomm, αsec], and the total weight is calculated as W'(e)=αcomp·Wcomp(e)+αcomm·Wcomm(e)+αsec·Wsec(e).

[0194] To capture the interaction effects between elements, an element interaction matrix I was designed. For any pair of elements (ei, ej), I(ei, ej) represents the additional cost or benefit incurred when using these two elements simultaneously. For example, certain cryptographic elements may produce a synergistic effect when used in conjunction with specific computational elements, reducing the overall cost; while other element combinations may lead to additional overhead. The interaction effects were established through experimental measurements and an expert knowledge base and were considered in the weighting calculations.

[0195] Ultimately, a complete element weight matrix was generated, containing the weight values ​​of approximately 120 protocol elements across three core dimensions, as well as the interaction coefficients between elements. This matrix serves as a crucial input for subsequent optimization algorithms, guiding the system to find the optimal protocol combination that balances performance and security.

[0196] For each element in the protocol element library, a multi-dimensional weight index is defined: computational cost (CPU cycles, memory usage, computational latency), communication overhead (transmission data volume, communication rounds, bandwidth requirements), and security strength (differential privacy parameter ε value, computational security, adversarial model strength). Through standardization and context weight adjustment, an element weight matrix is ​​constructed to provide a quantitative decision basis for subsequent optimization.

[0197] Step S34: Based on the element weight matrix, design and implement the minimum weight basis solution algorithm, select the protocol element combination with the differential privacy parameter ε value less than the preset security threshold and the minimum total computation cost and communication overhead from the protocol elements, and obtain the protocol minimum weight basis.

[0198] This step, based on the element weight matrix constructed in the previous steps, designs an efficient algorithm to find the minimum weight basis of the matroid. The minimum weight basis is the basis with the smallest total weight among all bases that satisfy all constraints. In the context of privacy-preserving computation protocols, the minimum weight basis corresponds to the combination of protocol elements that satisfies the security threshold requirement and minimizes the total cost.

[0199] First, the optimization objective is defined. Given a set of protocol elements E, an element weight matrix W, and a security threshold εmax, the objective is to find a subset B of E. E, such that: B is a basis of matroid M=(E,I) (i.e., B is a maximal independent set); the effective difference privacy parameter ε(B) for all combinations of elements of B is ≤εmax; the total weight W(B) = ∑e∈B W'(e) + ∑(ei,ej)∈B×BI(ei,ej) is minimized. Where W'(e) is the weighted total cost of element e, and I(ei,ej) is the interaction effect between elements.

[0200] To address this optimization problem, a solution algorithm combining a greedy strategy and local search was designed. The algorithm consists of four main stages:

[0201] The initialization phase sets up the necessary data structures and parameters for the algorithm. First, it determines the set of functions F that the protocol must support; each function f∈F can be implemented by multiple elements, and a function coverage mapping is constructed. :E→2 F Let represent the set of functions supported by each element. Then, the element weight matrix is ​​preprocessed, the effective weights after considering interactions are calculated, and the elements are initially screened according to the security strength, removing elements whose ε values ​​significantly exceed the security threshold.

[0202] The initial basis is constructed using a greedy strategy during the basic construction phase. The algorithm starts from the empty set S = ... Initially, elements are added incrementally until S becomes a basis. At each step, the algorithm selects an element e that satisfies three conditions: e satisfies the independence constraint with the current set S (i.e., ε(S∪{e})≤εmax and η(S∪{e})≥ηmin); e can cover some currently uncovered features; and among the elements satisfying the first two conditions, e has the smallest marginal weight increase ΔW(e,S)=W'(e)+∑e'∈SI(e,e'). The algorithm continues adding elements until set S covers all necessary features and no more elements can be added while maintaining independence.

[0203] In the local optimization phase, the total weight of the basis is further reduced through element substitution. For the current basis B, the algorithm considers each element b∈B and attempts to replace b with an element e that is not in B, provided that the resulting set (B\{b})∪{e} is still a basis and satisfies the safety threshold constraint. If the substitution reduces the total weight, the substitution is performed. This process is iterated until the weight cannot be reduced by simple substitution. To improve search efficiency, the algorithm uses a heuristic strategy to prioritize substitution attempts for element pairs with similar weights.

[0204] The global refinement phase employs more sophisticated optimization techniques to find a globally optimal solution. A stochastic search strategy based on simulated annealing is implemented, allowing temporary weight increases in substitutions to escape local optima. Simultaneously, multi-element substitution operations are attempted, replacing multiple elements at once to capture more complex optimization opportunities. To improve efficiency, the algorithm uses dynamic programming to pre-calculate the impact of certain substitution operations and employs a lazy evaluation strategy, calculating the full weight change only when necessary.

[0205] The algorithm has a time complexity of O(|E|²·|F|·log|E|), where |E| is the number of elements and |F| is the number of functions. In practical tests, for a protocol optimization problem containing approximately 120 elements and 30 functions, the algorithm can find a high-quality solution within a few seconds. Compared to exhaustive search, this algorithm significantly reduces computational complexity, and experimental results show that the weights of the solutions obtained are typically within 1.1 times that of the theoretical optimal solution.

[0206] This algorithm successfully identifies the combination of protocol elements that meets the security threshold requirements and minimizes the total cost, forming a minimum weighted basis for the protocol. This minimum weighted basis ensures that the protocol's security is no less than a preset standard, while minimizing computational costs and communication overhead, achieving an optimal balance between security and efficiency.

[0207] Based on the element weight matrix, a minimum weight basis solution algorithm combining greedy strategy and local search is designed: starting from an empty set, elements that satisfy security constraints and minimize marginal cost are gradually added to form an initial basis; local optimization is performed through single-element replacement and multi-element exchange; simulated annealing strategy is applied to explore a wider solution space, and finally the protocol element combination with the minimum total computational cost and communication overhead is obtained under the premise that the differential privacy parameter ε is less than the preset security threshold.

[0208] Step S35: Based on the minimum weight basis of the protocol, construct a protocol optimization framework to realize the dynamic combination and formal verification of protocol elements, and obtain the optimized privacy computing protocol framework.

[0209] This step, based on the minimum weight basis of the protocol obtained in the preceding steps, constructs a complete protocol optimization framework, realizing element combination, protocol generation, and formal verification, ultimately forming an optimized privacy computing protocol framework. This framework not only generates static protocol specifications but also supports dynamic combination and optimization at runtime.

[0210] First, a protocol synthesis engine was designed to transform abstract protocol elements into concrete protocol specifications and implementation code. Protocol synthesis involves three key steps:

[0211] Standardizing element interfaces ensures seamless integration of selected protocol elements. A unified element interface specification is defined, including input / output formats, state management, and error handling mechanisms. For elements with incompatible interfaces, necessary adapter layers or wrappers are generated to ensure interoperability between elements. Interface standardization uses a typed interface description language, explicitly defining data structures and function signatures, and supporting static type checking and automatic compatibility verification.

[0212] Protocol workflow orchestration organizes independent elements into an ordered execution flow. Based on the dependencies between elements and data flow requirements, a directed acyclic graph (DAG) for protocol execution is constructed. Nodes in the graph represent protocol elements, and edges represent data flow and control flow dependencies. Topological sorting is applied to determine the execution order of elements and to identify subsets of elements that can be executed in parallel to improve efficiency. Workflow orchestration also considers exception handling and recovery mechanisms, defining rollback and retry strategies for element failures.

[0213] Parameter configuration optimization determines the optimal parameter settings for selected elements. Based on statistical performance models and the characteristics of the target application scenario, key parameters for each element are adjusted. For example, for differential privacy elements, specific ε values ​​are assigned based on the overall privacy budget εmax and data sensitivity; for encryption elements, appropriate key lengths and encryption modes are selected based on security requirements and performance objectives. Parameter optimization employs a multi-objective optimization method to maximize performance while ensuring security thresholds are met.

[0214] Secondly, a formal verification framework for the protocol was implemented to ensure that the synthesized protocol meets the expected security and functional correctness requirements. The verification framework comprises three core components:

[0215] The security verification module checks whether the protocol satisfies predefined security properties. It uses symbolic execution and theorem proving techniques to verify the differential privacy guarantees of the protocol under security models (such as semi-honest or malicious adversary models). Security verification pays particular attention to the composite security properties resulting from combinations of elements, ensuring that combinations of individually secure elements do not introduce unexpected security vulnerabilities. Automated verification of the differential privacy composition theorem is implemented, ensuring that combinations of multiple differential privacy mechanisms do not exceed the total privacy budget.

[0216] Functional correctness verification ensures that the protocol can correctly achieve its intended functions. A combination of model checking and unit testing is used to verify the correctness of the protocol's output under various input conditions. Functional verification includes boundary condition testing, handling of abnormal inputs, and performance guarantee checks. A test suite for medical data computation was built, containing common data patterns and computational tasks from real-world scenarios, to comprehensively verify the protocol's functionality.

[0217] Resource compliance verification checks whether the protocol's resource requirements are within the constraints of the target environment. It analyzes the protocol's maximum memory usage, computational complexity, and network bandwidth requirements to ensure they meet preset resource limits. For protocols with resource requirements close to the limits, it provides resource usage optimization suggestions, such as adjusting batch processing strategies or optimizing memory usage.

[0218] Next, a dynamic protocol optimization mechanism was developed, supporting runtime adjustments to protocol configuration based on environmental changes. Dynamic optimization includes three main features:

[0219] Environment-aware adaptation adjusts protocol parameters and components based on runtime conditions. It monitors changes in computing resource utilization, network conditions, and data characteristics, triggering protocol reconfiguration when significant changes are detected. For example, when network bandwidth decreases, it may switch to data encoding elements with higher compression rates; when changes in data distribution are detected, it may adjust differential privacy noise parameters to maintain privacy protection.

[0220] Incremental protocol enhancement allows the protocol to be optimized gradually after initial deployment. Performance metrics and patterns of protocol execution are collected, and a predictive model is trained based on historical data to identify performance bottlenecks and optimization opportunities. Possible element replacements or parameter adjustments are periodically evaluated; if the predicted performance improvement exceeds a threshold, a protocol upgrade is recommended. Incremental enhancement employs A / B testing, deploying and validating optimization solutions on a small scale before full-scale rollout.

[0221] The security degradation mechanism handles protocol execution under extreme conditions. When resources are extremely limited or some components fail, the system can degrade protocol functions to ensure the availability of basic services while maintaining core security guarantees. Security degradation is based on predefined degradation paths, explicitly specifying which elements can be degraded and their minimum security requirements, ensuring that the degradation process is controlled and auditable.

[0222] Ultimately, a complete optimized privacy-preserving computation protocol framework was formed by integrating a protocol synthesis engine, a formal verification framework, and a dynamic optimization mechanism. This framework not only provides static protocol specifications but also supports dynamic assembly, verification, and runtime optimization of protocols, enabling secure, efficient, and adaptive processing of privacy-preserving computations for medical data.

[0223] Based on the minimum weight base of the protocol, a comprehensive protocol optimization framework is constructed: abstract elements are transformed into executable protocols through element interface standardization, workflow orchestration, and parameter configuration optimization; security verification, functional correctness verification, and resource compliance verification are implemented to ensure that the protocol meets the expected requirements; environment-aware adaptation, progressive protocol enhancement, and security degradation mechanisms are developed to support dynamic optimization at runtime, ultimately forming a privacy computing protocol framework that balances security, efficiency, and adaptability.

[0224] In one embodiment of the present invention, step S4 involves performing privacy sensitivity analysis and classification on medical data based on the optimized privacy computing protocol framework, designing a hybrid encryption strategy, and implementing a distributed key generation, distribution, and management system to obtain a medical data encryption and decryption system.

[0225] Specifically, in step S41, based on the optimized privacy computing protocol framework, a privacy sensitivity quantification analysis is performed on patient identity information, diagnostic records, and genetic data in medical data. The data is divided into three levels of sensitivity from high to low: high sensitivity, medium sensitivity, and low sensitivity. A multi-level protection strategy is established to obtain a data sensitivity grading model.

[0226] First, a Privacy Impact Assessment (PIA) methodology is employed to conduct a systematic privacy sensitivity analysis of various types of information within medical data. Privacy Impact Assessment is a structured approach used to identify, assess, and mitigate the potential impact of information systems on individual privacy. In the context of medical data, PIA specifically focuses on three dimensions: data re-identification risk, the degree of potential harm, and regulatory compliance requirements.

[0227] To address the unique characteristics of medical data, a quantitative scoring system for measuring medical data privacy sensitivity was developed. Data sensitivity is assessed across four key dimensions: Identification Risk (IR), which measures the likelihood of data being used to identify individuals; Harm Level (HD), which assesses the severity of potential harm from data breaches; Correlation (CO), which considers the increased privacy risks when the data is combined with other data; and Regulatory Requirement (RR), which reflects the stringency of applicable regulations regarding data protection. Each dimension is scored from 1 to 10, and a weighted sum is used to obtain the overall privacy sensitivity score for the data.

[0228] Based on the quantitative scoring results, a three-tiered data sensitivity classification model was established: High-sensitivity data (score > 7.5), including patient identification information (such as name, ID number, detailed address), genetic disease gene marker data, mental illness records, and HIV / AIDS treatment information. This type of data is directly linked to personal identity, and its disclosure may cause serious discrimination or harm, requiring the highest level of protection; Medium-sensitivity data (score 4-7.5), including general diagnostic records, treatment plans, prescription information, and routine test results. Although this type of data does not directly identify individuals, it contains important health information, and its disclosure would constitute a privacy violation, requiring a medium level of protection; Low-sensitivity data (score < 4), including anonymized statistical information, de-identified research data, and aggregated health trends. This type of data has already been anonymized, and the risk of re-identification is low, allowing for basic level protection.

[0229] For boundary data that is difficult to classify clearly, this step adopts a risk-first principle, classifying it into a higher sensitivity level. A dynamic sensitivity assessment mechanism is also established to adjust the sensitivity level based on the data combination usage scenario and changes in the external environment, ensuring the adaptability and rationality of the classification.

[0230] By quantitatively analyzing the privacy sensitivity of different types of information in medical data, a three-tiered protection model was established: highly sensitive data includes patient identity information and genetic disease gene data; moderately sensitive data includes diagnostic records and treatment plans; and low-sensitive data includes anonymized statistical information and research data, providing a basis for subsequent differentiated encryption strategies.

[0231] Step S42: Based on the data sensitivity classification model, design a hybrid encryption strategy that combines symmetric and asymmetric encryption. Select encryption schemes of different strengths for data with different sensitivities to obtain the hybrid encryption strategy.

[0232] This step, based on the sensitivity grading model established in the preceding steps, designs a multi-layered hybrid encryption strategy to provide differentiated protection for medical data with different sensitivity levels. The hybrid encryption strategy combines multiple encryption technologies, fully utilizing the advantages of various encryption algorithms to balance security, performance, and functionality requirements.

[0233] For highly sensitive data, this strategy employs a multi-layered defense design, combining asymmetric encryption, homomorphic encryption, and zero-knowledge proof techniques. The core architecture includes: a base layer using RSA-4096 or Elliptic Curve Cryptography (ECC) to provide strong key protection; a functional layer using Fully Homomorphic Encryption (FHE) or Partially Homomorphic Encryption (PHE) to support ciphertext computation; and a verification layer integrating zero-knowledge proofs to ensure computational correctness without revealing the original data. The encryption scheme for highly sensitive data places particular emphasis on security, prioritizing absolute data security even at the cost of higher computational costs.

[0234] For moderately sensitive data, this strategy employs a balanced protection design, combining efficient symmetric encryption with secure key management. Its architecture includes: a data encryption layer using the AES-256 algorithm to efficiently encrypt large amounts of medical data; a key protection layer employing distributed key management to ensure the security of encryption keys; and an access control layer implementing attribute-based access control (ABAC) to ensure that only authorized users can access specific data. This encryption scheme for moderately sensitive data strikes a balance between security and efficiency, making it suitable for most routine medical data processing scenarios.

[0235] For low-sensitivity data, this strategy employs a lightweight protection design, combining lightweight encryption with differential privacy techniques. Its architecture includes: a basic protection layer using lightweight encryption algorithms such as ChaCha20 to provide basic confidentiality; a privacy protection layer applying differential privacy techniques to add calibration noise to the data to prevent statistical inference attacks; and a data transformation layer using format-preserving encryption (FPE) to maintain the data format while providing encryption protection. The encryption scheme for low-sensitivity data prioritizes efficiency and usability, minimizing performance overhead while providing necessary security guarantees.

[0236] The key innovation in the hybrid encryption strategy is the "adaptive security level switching" mechanism. This mechanism allows for automatic adjustment of the encryption scheme at different stages of the data lifecycle, such as automatically upgrading the security level in high-risk environments or temporarily switching to a more efficient encryption scheme in computationally intensive scenarios. Secure re-encryption technology is used during the switching process, allowing for encryption mode conversion without complete decryption, thus minimizing security risks.

[0237] Based on the data sensitivity classification model, a hybrid encryption strategy was designed: high-strength asymmetric encryption combined with homomorphic encryption was used for highly sensitive data; a scheme combining symmetric encryption and distributed key management was used for medium-sensitive data; and lightweight encryption or differential privacy technology was used for low-sensitive data, thus achieving a balance between security and efficiency.

[0238] Step S43: Based on the hybrid encryption strategy and the computational cost constraints in the optimized privacy computing protocol framework, optimize the key expansion rounds and block mode parameters of the encryption algorithm to obtain the optimized encryption algorithm set.

[0239] Based on the hybrid encryption strategy designed in the preceding steps, and combined with the computational cost constraints provided by the protocol framework in Example 3, this step refines the encryption algorithm parameters to construct a set of encryption algorithms suitable for the characteristics of medical data.

[0240] First, this step develops a data feature-aware parameter optimization strategy tailored to the characteristics of different types of medical data. Medical data encompasses various formats, such as structured electronic health records (EHRs), semi-structured medical images, and unstructured clinical notes, each with unique data characteristics and processing requirements.

[0241] For medical image data, this step optimizes the number of key expansion rounds in the symmetric encryption algorithm. Key expansion is the process in symmetric encryption algorithms that expands the master key into multiple rounds of subkeys; more rounds result in higher security but slower speed. Considering the large volume of medical image data and the high processing speed requirements, for moderately sensitive image data, the number of key expansion rounds in the standard AES-256 algorithm is optimized from 14 rounds to 10 rounds. Simultaneously, a pre-computation technique is used to cache the round keys, reducing real-time computation overhead. Experiments show that this optimization improves encryption processing speed by up to 40% while maintaining sufficient security strength, significantly improving the processing performance of large medical image datasets.

[0242] For genomic data, this step optimized the working mode parameters of the block cipher. The block mode determines how data larger than a single block is processed, and different modes differ in security, error propagation, and parallel processing capabilities. For highly sensitive genomic data, the Galois / Counter Mode (GCM) was selected, and the authentication tag length and initialization vector generation strategy were optimized. Specifically, a 128-bit authentication tag is used for complete genomic data to ensure maximum integrity, while a 64-bit authentication tag is used for frequently accessed gene fragments to balance security and performance. Furthermore, through algorithmic improvements, the GCM mode supports streaming processing of ultra-large genomic datasets, solving the memory limitations of traditional implementations when processing terabyte-scale data.

[0243] For clinical text data, this step optimizes the parameters of a lightweight encryption algorithm. For low-to-medium sensitivity de-identified clinical text, the ChaCha20-Poly1305 algorithm is adopted, and the round function parameters are optimized specifically for the characteristics of medical text. Traditional ChaCha20 uses a 20-round function; this solution, based on security analysis, reduces this to 12 rounds for low-sensitivity text data, improving processing speed by approximately 35% while maintaining a sufficient security margin.

[0244] For medical data requiring homomorphic computation, this step optimizes the parameters of the homomorphic encryption scheme. Homomorphic encryption allows direct computation on ciphertext, but parameter selection directly impacts performance and accuracy. For moderately sensitive data requiring statistical analysis, the polynomial modulus size and coefficient modulus chain of the CKKS scheme are optimized to minimize computational overhead while maintaining accuracy. Through data feature analysis, precision parameters are customized for clinical indicator analysis scenarios, improving homomorphic computation performance by nearly 50% while maintaining the medical validity of the analysis results.

[0245] Through the parameter optimizations described above, this step constructs a complete set of optimized encryption algorithms, providing a secure and efficient encryption foundation for privacy-preserving computations of medical data. Experimental results show that the optimized algorithm set improves processing efficiency by an average of 30-45% compared to standard parameter configurations in typical medical data processing scenarios, while maintaining the protection strength required to meet medical data security requirements.

[0246] The encryption algorithm parameters were optimized to suit the characteristics of medical data. For example, fewer key expansion rounds were used for medical image data with high processing speed requirements, and a more complex grouping mode was used for genetic data with extremely high security requirements. Through fine-tuning, the performance of the encryption algorithm was optimized, ensuring that computational efficiency was maximized while meeting security requirements.

[0247] Step S44: Based on the optimized encryption algorithm set, design and implement a distributed key generation, distribution and management system to ensure key security in a multi-organization collaborative environment, thus obtaining the key management system.

[0248] This step, based on the optimized encryption algorithm set from the preceding steps, designs and implements a complete distributed key management system to address the key security challenges in multi-institutional medical data collaboration. Distributed key management is a technology that distributes key generation, storage, and use across multiple entities, avoiding single points of failure and single points of trust issues.

[0249] First, this step designs a distributed key generation mechanism based on threshold cryptography. Threshold cryptography is a technique that divides a secret into multiple parts, requiring a predetermined threshold number of parts to reconstruct the key. In multi-institutional medical collaborations, a single institution holding the complete key poses security risks and trust issues. This system adopts an (t,n) threshold scheme, where n is the total number of participating institutions, and t is the minimum number of institutions required to reconstruct the key, typically set to at least 2 / 3 of n, balancing security and availability.

[0250] For highly sensitive data, a distributed key generation protocol based on multi-party secure computation was implemented. This protocol allows all participants to jointly generate encryption keys without any single party obtaining the complete key. Specifically, the system uses Pedersen VSS to achieve verifiable secret sharing, ensuring that each participant correctly executes the protocol; it combines elliptic curve cryptography (ECC) to reduce communication overhead, enabling efficient key generation even in wide area network environments; and it implements distributed random number generation to ensure that the randomness of the key is not controlled by any single party. This design guarantees that even if t-1 institutions are compromised, the complete key will not be leaked.

[0251] For the key distribution stage, a secure credential exchange mechanism was designed in this step. This mechanism uses a combination of one-time passwords and out-of-band authentication to ensure that key-sharing information is only sent to legitimate recipients. The system assigns a unique identifier and certificate to each organization, building a PKI-based trust framework; it uses TLS 1.3 to provide communication encryption and adds additional session binding to ensure communication integrity; it adopts a hierarchical key structure, using a master key to derive session keys, limiting the use and lifespan of a single key and reducing the impact of leakage.

[0252] In terms of key management, complete lifecycle management has been implemented. Key lifecycle management is a systematic method for managing the entire process of keys from generation to destruction, which is crucial for ensuring long-term security. Specifically, it includes: an automated key rotation mechanism that regularly updates all encryption keys to reduce the risk of using the same key for an extended period; a hierarchical key architecture that divides keys into root keys, intermediate keys, and operational keys, with different update frequencies for each level; secure key backup using threshold encryption to store key backups and prevent backups from becoming security weaknesses; and key usage auditing that records all key operations and reviews them regularly to promptly identify abnormal usage patterns.

[0253] To address emergencies, a key revocation and recovery mechanism was designed. When a key leak is detected or a participating institution withdraws, the system can quickly revoke the relevant key and notify all participants. The revocation process employs a two-phase commit protocol to ensure all institutions synchronously update their key status; an emergency key regeneration process is implemented to quickly create a new key after revocation, minimizing service interruption; and key history management is supported, allowing access to data encrypted with older keys when necessary, preventing data loss.

[0254] Through the above design, this step implemented a complete distributed key management system, providing robust key security for multi-institutional medical data collaboration. The system performed excellently in actual deployment tests at five hospitals, supporting secure cross-institutional medical data analysis, while key management overhead accounted for less than 5% of total computing resources, verifying its practicality and efficiency.

[0255] A distributed key management system supporting multi-institutional collaborative environments has been implemented. A threshold key sharing scheme is adopted to ensure that the complete key will not be leaked even if some institutions are compromised. Core functions such as secure key generation, controllable distribution, regular updates and emergency revocation have been implemented, providing key security for collaborative analysis of medical data from multiple institutions.

[0256] Step S45: Based on the key management system, the encryption and decryption functions are encapsulated into a unified interface and service, supporting calls from various privacy computing modules, to obtain the medical data encryption and decryption system.

[0257] This step, building upon the key management system implemented in the preceding steps, develops a unified encryption / decryption service interface layer. This achieves seamless integration between various components, forming a complete medical data encryption / decryption system. The unified interface layer serves as a bridge connecting the underlying encryption functions with the upper-layer applications; a well-designed interface is crucial for system usability and scalability.

[0258] First, this step defines a standardized encryption / decryption service interface specification. The interface specification adopts a Service-Oriented Architecture (SOA) design philosophy, encapsulating all encryption / decryption functions as independent services, supporting loosely coupled calls. The interface design follows four core principles: consistency principle, ensuring all interfaces maintain a consistent parameter structure and error handling mechanism; least privilege principle, exposing only necessary functions and hiding implementation details; composability principle, allowing basic interfaces to be combined to build complex functionalities; and scalability principle, supporting seamless addition of new algorithms and functions in the future.

[0259] The core interface set includes five types of functional interfaces: a basic encryption / decryption interface, supporting direct calls to various encryption algorithms, including symmetric encryption, asymmetric encryption, and hash function operations; a ciphertext calculation interface, supporting ciphertext addition, multiplication, and other operations under homomorphic encryption, facilitating data analysis in encrypted states; a key operation interface, supporting key generation, acquisition, and management functions, and connecting to a distributed key management system; a data conversion interface, supporting conversion between different encryption formats, such as converting standard encryption to homomorphic encryption; and a policy control interface, supporting automatic selection of encryption schemes based on data sensitivity.

[0260] This step implements the Intelligent Encryption Agent (IEA) component, which acts as an intermediary layer between the application and the encryption service. The IEA automatically identifies data types and sensitivity levels, selects the most suitable encryption scheme, and automates the encryption process. Its key functions include: a data classification engine that automatically identifies medical data types and sensitivity levels; a scheme selector that automatically selects the optimal encryption scheme based on data characteristics, security requirements, and performance requirements; a resource scheduler that monitors system resource usage and optimizes encryption task execution; and an adapter converter that handles differences in data formats and encryption protocols between different systems.

[0261] To support data operations in encrypted mode, this step developed the CiphertextOperation Library (COL). This library provides the ability to perform common medical data analysis operations in encrypted mode, which is key to achieving "usable but invisible." Core functionalities include: encrypted statistical analysis, supporting the calculation of statistics such as mean, variance, and correlation coefficient in encrypted mode; encrypted model training, supporting the training of simple machine learning models on encrypted data; encrypted data concatenation, securely merging encrypted medical data from different sources; and secure query execution, supporting conditional queries on encrypted databases.

[0262] This step also implements a complete system monitoring and auditing mechanism to ensure the secure operation of the encryption / decryption system. The monitoring component tracks the execution status, performance indicators, and error conditions of encryption / decryption operations in real time, supporting anomaly detection and automatic alarms. The auditing component records detailed logs of all encryption / decryption operations, including operation type, data identifier, algorithm used, and initiator information, supporting post-event review and compliance checks.

[0263] Finally, this step establishes an integration interface with existing healthcare systems, supporting seamless interoperability between the encryption / decryption system and Hospital Information Systems (HIS), Electronic Medical Records (EMR), and Health Information Exchange (HIE) platforms. The integration utilizes standardized medical data exchange protocols (such as HL7 FHIR) and employs a dedicated connector to simplify the integration process.

[0264] Through the design and implementation described above, this step completed the construction of a medical data encryption and decryption system, providing comprehensive protection for the secure sharing and analysis of medical data. In actual medical environment testing, it demonstrated good adaptability and performance, supporting diverse scenarios ranging from simple secure data storage to complex multi-party privacy computations.

[0265] The encryption and decryption functions are encapsulated into a unified API interface and service, providing standardized function calls for data encryption, ciphertext calculation, and result decryption. It supports seamless access for various privacy computing modules, achieving loose coupling between the encryption and decryption system and other modules, thus forming a complete medical data encryption and decryption system.

[0266] In one embodiment of the present invention, step S5 involves standardizing and encapsulating various privacy computing technologies and defining a unified interface based on the medical data encryption and decryption system, constructing an execution engine to achieve multi-technology collaborative scheduling and execution, and obtaining a medical data privacy computing fusion system.

[0267] Specifically, in step S51, based on the API interface provided by the medical data encryption and decryption system, federated learning, homomorphic encryption, and secure multi-party computation are standardized and encapsulated. The data input and output of each technology are connected to the encryption and decryption service of the medical data encryption and decryption system. A unified interface is defined, including a data encryption interface, a ciphertext computation interface, and a result decryption interface, to obtain a standardized privacy computation module set.

[0268] Based on the API interface provided by the medical data encryption / decryption system in the aforementioned embodiments, mainstream privacy computing technologies are standardized and encapsulated to construct a unified technology invocation framework. Standardized encapsulation refers to hiding the specific implementation details of different technologies behind a consistent interface, so that external callers do not need to care about the internal implementation differences and can call different technologies in a unified way.

[0269] First, a Technology Abstraction Layer (TAL) is implemented. This is an intermediate layer architecture used to bridge the native interfaces of different privacy computing technologies with a unified standard interface. TAL adopts the adapter design pattern, creating a dedicated adapter for each technology to convert technology-specific APIs into standardized interfaces. For federated learning, it encapsulates two modes: horizontal federated learning (scenarios with the same data features but different samples) and vertical federated learning (scenarios with different features but the same sample ID), and implements standard interfaces for core functions such as model initialization, parameter exchange, gradient aggregation, and model evaluation. For homomorphic encryption, it encapsulates schemes including partial homomorphic encryption (such as Paillier) and fully homomorphic encryption (such as CKKS), and implements a unified calling method for operations such as key generation, data encryption, homomorphic operations, and result decryption. For secure multi-party computation, it encapsulates implementations based on secret sharing, obfuscated circuits, and unintentional transmission, and provides standardized access to functions such as secure computation, comparison, and set operations.

[0270] To address the differences in data input and output formats across various technologies, a Data Conversion Adapter (DCA) was designed. DCA enables seamless data format conversion between different technologies, supporting the conversion of structured medical data (such as electronic health records), semi-structured data (such as medical reports), and unstructured data (such as medical images) across different technologies. The core mechanism is a Universal Medical Data Exchange Format (UMDEF), an intermediate representation format. All data input is first converted to UMDEF, and then to the format required by the target technology. This design significantly simplifies integration complexity, reducing the number of converters from n² (pairwise conversions between n technologies) to only 2n converters (bidirectional conversion from each technology to UMDEF).

[0271] This step defines three types of core unified interfaces, which constitute the standard calling protocol for privacy computing technology:

[0272] 1. The Data Encryption Interface is responsible for converting raw medical data into an encrypted form suitable for privacy-preserving computations. The interface design follows the "data type awareness" principle, automatically selecting the most appropriate encryption scheme based on the characteristics of the medical data. The interface includes three core methods: the `prepareData` method handles data preprocessing, normalization, and format conversion; the `encryptData` method performs the actual encryption operation, calling the corresponding functions of the medical data encryption / decryption system; and the `verifyEncryption` method verifies the correctness and integrity of the encryption result, ensuring the reliability of the subsequent computational foundation.

[0273] 2. The Ciphertext Computation Interface defines a standard method for performing computations on encrypted data. The interface employs a "polymorphic operation type" design, allowing the same method to perform different ciphertext operations depending on the type of input parameters. Core methods include: `computeStatistics` performs statistical analysis operations such as mean and variance calculations; `trainModel` implements model training, supporting various algorithms such as linear models and decision trees; `evaluateModel` performs model evaluation and validation; and `queryData` supports query operations on encrypted data. Internally, these methods automatically select the most suitable privacy-preserving computation technology based on data characteristics and task requirements.

[0274] 3. The Result Decryption Interface is responsible for securely converting computation results from ciphertext to plaintext. This interface implements an access-controlled decryption mechanism, ensuring that only authorized users can access the decrypted result. Core methods include: `authorizeDecryption` verifies the validity of the decryption request; `decryptResult` performs the result decryption operation; `auditDecryption` logs the decryption operation for subsequent auditing; and `transformResult` converts the decrypted result into the format required by the application.

[0275] Through the standardized encapsulation and interface definition described above, a unified framework for calling different privacy computing technologies is achieved, enabling application developers to easily use various privacy computing functions without needing to understand the underlying technical details. This design significantly reduces integration complexity and improves system scalability and maintainability.

[0276] The mainstream privacy computing technologies have been standardized and encapsulated, giving each technology module a unified interface specification, including data input interface, computing interface and result output interface, which enables seamless integration with encryption and decryption systems and forms a composable standardized privacy computing module set.

[0277] Step S52: Based on the standardized privacy computing module set, implement a parsing system for medical data analysis and model training tasks. Receive user-submitted requests for disease prediction model training, multi-center clinical data statistical analysis, and genomic data association analysis. Call the medical data encryption / decryption system to encrypt the original medical data. Decompose the requests for disease prediction model training, multi-center clinical data statistical analysis, and genomic data association analysis into a sequence of privacy computing sub-tasks to obtain a task execution plan.

[0278] Based on the standardized privacy-preserving computation module set constructed through the aforementioned steps, a medical data analysis task parsing system was developed, enabling intelligent conversion from high-level medical analysis needs to specific privacy-preserving computation tasks. The task parsing system acts as a bridge connecting user needs with underlying technological implementations, understanding analytical requirements within the medical context and transforming them into executable computational tasks.

[0279] First, a domain-specific language (MAL) was designed as the language for users to express their analytical needs. MAL is a declarative language specifically designed for medical data analysis scenarios, allowing medical professionals to describe their analytical requirements in a near-natural language manner. MAL's core features include: support for medical terminology, with built-in vocabulary of commonly used medical indicators and analytical methods; data source description, supporting the specification of multi-center data collaboration models; privacy constraint statements, allowing explicit specification of data sharing restrictions and security requirements; and expression of expected results, supporting the description of desired output formats and accuracy requirements. For example, users can use MAL to express complex needs such as "Based on patient data from three hospitals, train a type 2 diabetes prediction model without sharing original records, with a model accuracy requirement of AUC > 0.8."

[0280] The core of the task parsing system is the Requirement Comprehension Engine (RCE), which is responsible for parsing the analytical requirements expressed in the MAL and understanding their medical context and computational intent. RCE employs a two-stage parsing strategy: in the semantic analysis stage, it uses a medical ontology library and natural language processing techniques to identify medical concepts, data types, and operational intents within the requirements; in the computational mapping stage, it maps the identified semantic elements to specific computational tasks and data requirements. For example, when identifying "type 2 diabetes prediction," the system can automatically associate relevant risk factors (such as BMI, age, blood glucose, etc.) as required features and determine the appropriate model type (such as logistic regression or random forest).

[0281] To address task decomposition, a task planner was implemented, breaking down the overall analysis requirements into a series of sub-tasks. The task planner uses a directed acyclic graph (DAG) to represent task dependencies, where each node represents a sub-task and edges represent data flow or control dependencies. The planning process consists of three steps: First, identifying core computational tasks, such as feature engineering, model training, or statistical analysis; second, adding necessary pre- and post-processing steps for each core task, such as data encryption, format conversion, and result verification; and finally, optimizing the task graph, merging tasks that can be executed in parallel, and eliminating redundant operations. The task planner also implements Privacy-Aware Task Reordering (PATR) technology, an optimization strategy that adjusts the task execution order to minimize the exposure scope and duration of sensitive data, further enhancing privacy protection.

[0282] For three typical medical analysis needs, specialized task templates and parsing logic have been implemented:

[0283] For disease prediction model training needs, it can automatically identify the prediction target (such as a specific disease) and potential features, and generate an execution plan that includes four stages: data preparation, feature selection, model training, and validation evaluation. It will automatically select a suitable federated learning scheme based on data distribution characteristics (such as horizontal or vertical distribution) and privacy requirements. For example, for horizontally distributed patient data, it will select horizontal federated learning and configure an appropriate aggregation strategy; for scenarios with vertically distributed features, it will select vertical federated learning and design a secure entity alignment scheme.

[0284] For the statistical analysis needs of multi-center clinical data, it can identify target statistical indicators (such as correlation, hypothesis testing, etc.) and the required data types, and generate an execution plan that includes data encryption, secure statistical computation, and result validation. Based on the type of computation and accuracy requirements, it intelligently selects either homomorphic encryption or secure multi-party computation techniques. For example, for simple descriptive statistics (such as mean and variance), it will prioritize partially homomorphic encryption, which has higher computational efficiency; while for complex hypothesis testing, it may choose a secure multi-party computation scheme based on obfuscated circuits.

[0285] For genomic data association analysis needs, we understand the specific requirements of genotype-phenotype association analysis and design execution plans that include data encryption, secure association testing, and multiple validation corrections. Considering the high sensitivity and large scale of genomic data, we prioritize technologies that provide high security guarantees and design specialized data partitioning strategies to handle large-scale genomic data.

[0286] The task execution plan is generated using Resource-Aware Planning (RAP), a planning strategy that takes into account the resource constraints of the computing environment. RAP dynamically adjusts the granularity of task decomposition and resource allocation based on currently available computing resources, network bandwidth, and the computing capabilities of the participants. For example, in resource-constrained environments, it increases task merging and reduces intermediate data transfer; while in high-performance environments, it increases parallelism to improve execution speed.

[0287] An intelligent task parsing system was developed that can receive and understand high-level medical data analysis needs, such as disease prediction model training or multi-center clinical data statistical analysis, and decompose them into a series of privacy-preserving computation sub-tasks to form an executable task plan, including steps such as data preparation, feature processing, model training and result verification.

[0288] Step S53: Based on the task execution plan, construct the privacy computing execution engine core, integrate the key management function of the medical data encryption and decryption system, realize multi-technology collaborative scheduling and execution, support data operations in encrypted state, and obtain the execution engine core system.

[0289] Based on the task execution plan generated by the aforementioned steps, a core system for a privacy-preserving computation execution engine was constructed to achieve task scheduling and collaborative execution of multiple technologies. The execution engine is the core component of the integrated system, responsible for coordinating various privacy-preserving computation technology modules to ensure the secure and efficient completion of tasks.

[0290] First, a multi-tier execution architecture (MEA) was designed, dividing the execution engine into three layers: a scheduling layer, a coordination layer, and an execution layer. The scheduling layer is responsible for global task management, receiving task execution plans and allocating resources; the coordination layer is responsible for the collaboration and interaction between multiple technologies, managing data flow and intermediate results; and the execution layer is responsible for the invocation and execution of specific privacy-preserving computation technologies. This layered design achieves separation of concerns, giving the system good modularity and scalability.

[0291] The core of the execution engine is the Technology Scheduler, responsible for selecting the most suitable privacy-preserving computation technology for each subtask and coordinating its execution. The scheduler implements a Context-Aware Technology Selection (CATS) mechanism, a method that dynamically selects the best technology implementation based on task characteristics, data features, and the runtime environment. CATS considers four key factors: data characteristics, such as scale, distribution, and sensitivity; computational characteristics, such as algorithm complexity and parallelization potential; security requirements, such as the strength of privacy protection and attack models; and resource constraints, such as computing power and network bandwidth. By comprehensively evaluating these factors, CATS can find the optimal technology solution that balances performance and security for each subtask.

[0292] To support seamless collaboration between different technologies, a Technology Coordination Layer (TCL) was implemented. TCL handles data exchange and result integration between different privacy-preserving computing technologies, ensuring the consistency and security of the entire computing process. Its core component is the Secure Intermediate State (SIS), a special data representation that guarantees intermediate results remain encrypted throughout technology switching. For example, when the workflow switches from the federated learning phase to the homomorphic encryption phase, TCL securely converts the federated learning output to a homomorphic encrypted input format without exposing the original data. This design ensures the "end-to-end encryption" principle, meaning that data remains encrypted throughout the entire processing.

[0293] This step is tightly integrated with the medical data encryption / decryption system, enabling coordinated key lifecycle management and execution processes. This integration employs an Execution Phase Key Synchronization (EPKS) ​​mechanism, automatically coordinating key generation, use, and destruction based on the task execution phase. EPKS maintains independent key contexts for different execution phases, ensuring that even if a security issue occurs in a single phase, it will not affect overall data security. For example, the keys used in the model training phase are independent of those used in the result verification phase, limiting the scope of potential attacks.

[0294] To support complex medical analysis scenarios, this step implements the Advanced Ciphertext Operation Engine (ACOE), providing a series of data manipulation functions that execute in an encrypted state. ACOE comprises four functional modules: a encrypted statistics module, supporting the calculation of statistics such as mean, variance, and correlation coefficient; a encrypted machine learning module, supporting the training of common models such as linear regression and decision trees; a encrypted set operation module, supporting set operations such as intersection and union; and a encrypted condition processing module, supporting conditional filtering without revealing the condition values. These modules are implemented using an "algorithm-technology mapping" approach, meaning that each operation has multiple technical implementations, and the system selects the most appropriate version based on the context.

[0295] It also implements a Fault-Tolerant Execution Mechanism (FTEM) to ensure computational stability in a distributed environment. FTEM includes three key functions: State Checkpoint, which periodically saves intermediate states to support fault recovery; Anomaly Detection, which monitors execution anomalies and triggers corresponding countermeasures; and Recovery Strategy, which defines recovery schemes for different anomaly scenarios. For example, when a participant's network is interrupted, the system can pause execution and save its state, resuming from the breakpoint once the connection is restored, without needing to restart the entire computation process.

[0296] Through the above design and implementation, this step constructs a powerful and flexible core system for a privacy-preserving computation execution engine. This system can intelligently schedule different privacy-preserving computation technologies according to task requirements, enabling multi-technology collaborative execution and ensuring that sensitive medical data remains encrypted and protected throughout the entire computation process, achieving a truly "usable but invisible" data processing paradigm.

[0297] An execution engine core system was built, integrating the key management function of the encryption and decryption system. It realizes the coordinated scheduling and execution of multiple privacy computing technologies, supports data operations and model training in encrypted state, and ensures that the data is always in an encrypted protection state during the computing process, realizing the "usable but invisible" nature of medical data.

[0298] Step S54: Based on the core system of the execution engine, implement a computing resource monitoring and performance analysis system to monitor the encryption and decryption overhead and key management status of the medical data encryption and decryption system, support runtime technology switching and parameter tuning, and obtain a self-optimizing execution environment.

[0299] Based on the core execution engine system built using the aforementioned steps, a computing resource monitoring and performance analysis system was developed, realizing a self-optimizing execution environment. Resource monitoring and performance analysis are key components for improving the efficiency and reliability of privacy-preserving computing, enabling real-time assessment of system status and triggering optimization measures.

[0300] First, this step designs a Hierarchical Monitoring Architecture (HMA) to monitor the system status comprehensively from the infrastructure layer to the application layer. The infrastructure layer monitors hardware resource usage, such as CPU, memory, storage, and network bandwidth; the middleware layer monitors encryption operation performance and key management status, such as encryption / decryption time, key usage frequency, and expiration date; and the application layer monitors task execution metrics, such as subtask completion time, intermediate result size, and final accuracy. This layered design provides a comprehensive view from underlying resources to high-level applications, supporting multi-dimensional performance evaluation and optimization.

[0301] At the heart of the monitoring system is the Privacy-preserving Computation Performance Analysis Engine (PCPAE), responsible for collecting and analyzing various performance metrics. PCPAE implements Low-interference Monitoring (LIM) technology, a specially designed performance data collection method that obtains detailed performance metrics while minimizing the impact on normal computation. LIM uses a sampling strategy rather than continuous monitoring and performs more intensive data collection during idle periods, ensuring that monitoring itself does not become a performance bottleneck. PCPAE generates three key analyses: resource utilization analysis, identifying resource usage bottlenecks and optimization opportunities; performance trend analysis, tracking metric changes over time and predicting potential performance problems; and benchmark analysis, comparing current performance with historical records and theoretical optimal values ​​to quantify the optimization space.

[0302] For medical data encryption and decryption systems, a dedicated Encryption Performance Monitoring Module (EPMM) has been implemented. EPMM tracks four key metrics: encryption / decryption latency, recording encryption and decryption times for different data types and sizes; throughput, measuring the amount of encrypted data processed per unit time; key operation overhead, monitoring the resource consumption of key generation, distribution, and updates; and memory usage patterns, analyzing the memory usage characteristics of encryption operations. Through these metrics, the system can accurately assess the performance cost of encryption operations, guiding subsequent optimizations.

[0303] This step innovatively designs a Dynamic Technology Switching Mechanism (DTSM), supporting adjustments to technology selection based on performance feedback during task execution. DTSM operates based on a Threshold Triggering Strategy (TTS), automatically triggering technology re-evaluation when performance metrics exceed preset thresholds. For example, if the execution time of a homomorphic encryption operation exceeds 150% of the expected time, the system may switch to a more efficient approximate computation method; or when network bandwidth drops below a certain threshold, it may switch from communication-intensive secure multi-party computation to a more bandwidth-efficient federated learning scheme. During technology switching, DTSM ensures secure data transfer and maintains continuity of privacy protection.

[0304] To achieve refined performance optimization, a Parameter Auto-Tuning System (PATS) was developed. PATS uses an online learning algorithm to automatically adjust key parameters of privacy-preserving computation techniques based on real-time performance feedback. The tuning process employs Bayesian optimization methods, striking a balance between exploring new parameter combinations and utilizing known good configurations, enabling it to find near-optimal parameter settings with fewer attempts. For federated learning, PATS can dynamically adjust batch size, aggregation frequency, and the number of local iterations; for homomorphic encryption, it can optimize the multinomial modulus, noise parameters, and multiplication depth; and for secure multi-party computation, it can adjust the circuit optimization level, parallelism, and communication batch size.

[0305] It also implements a Resource Adaptive Allocation Mechanism (RAAM), which dynamically adjusts resource allocation based on task importance and resource requirements. RAAM maintains a resource priority queue and allocates computing resources according to the urgency and importance of tasks. For critical medical analysis tasks (such as emergency patient data analysis), the system allocates more resources and prioritizes scheduling; while for routine research analyses, execution may be scheduled only when resources are sufficient. This dynamic resource allocation strategy ensures the system's responsiveness and fairness when facing multi-task loads.

[0306] By integrating the aforementioned components, this step constructs a self-optimizing execution environment capable of automatically adjusting execution strategies based on runtime status and performance feedback, continuously optimizing system performance. The self-optimizing environment employs a "closed-loop feedback control" model, forming a continuous optimization cycle of monitoring, analysis, adjustment, and verification. This design enables the system to adapt to different medical data processing scenarios and changing computing environments, maintaining consistently high operational efficiency.

[0307] A resource monitoring and performance analysis system was developed to monitor computing resource usage, encryption and decryption operation overhead, and key management status in real time. It supports dynamic adjustment of computing parameters and switching of technologies during runtime, and continuously optimizes execution efficiency based on performance feedback, forming an optimized execution environment with adaptive capabilities.

[0308] Step S55: Based on the self-optimizing execution environment, design and implement the application programming interface and result visualization display system, call the medical data encryption and decryption system to decrypt the calculation results and display them, thus obtaining the medical data privacy calculation fusion system.

[0309] Based on the self-optimizing execution environment built using the aforementioned steps, an application interface and result display system for end users were designed and implemented, completing the overall construction of the medical data privacy computing fusion system. The application layer interface and visualization are key components of the system's interaction with users, determining the system's usability and user experience.

[0310] First, a Layered Application Programming Interface (LAPI) was designed to provide suitable access methods for users with different technical backgrounds. The LAPI comprises three layers: a high-level API for clinicians and medical researchers, providing simple and intuitive function calls such as `predictDisease()` and `analyzeCorrelation()`, without requiring knowledge of underlying technical details; a mid-level API for bioinformaticians and medical data analysts, offering more flexible analysis control and allowing for customized analysis workflows and parameters; and a low-level API for technical experts, providing direct access to underlying privacy-preserving computation functions and supporting highly customized analysis schemes. This layered design enables users with different backgrounds to effectively use the system, lowering the adoption threshold.

[0311] To facilitate integration with existing healthcare systems, this step implements Standard Medical Interface Adapters (SMIAs). SMIAs support interfacing with mainstream healthcare information systems and standards, including HL7 FHIR (for medical data exchange), DICOM (for medical imaging), and OMOP (for observational medical data). The adapters handle data model mapping, format conversion, and semantic matching, enabling the converged system to seamlessly integrate into the existing IT ecosystem of healthcare institutions. This design reduces deployment barriers and accelerates the system's application in real-world healthcare environments.

[0312] To ensure secure presentation of computational results, a Result Decryption and Authorized Presentation Mechanism (RDAPM) was designed. RDAPM implements role- and purpose-based access control, ensuring that only authorized users can view the decrypted results, and that access is limited to the minimum scope required by their responsibilities. Key functions include: tiered result presentation, displaying results of different granularities and sensitivities based on user permissions; purpose-linked access, associating result access with specific medical or research purposes; access audit trails, recording all result access activities to support compliance audits; and differential privacy filtering, applying differential privacy processing to highly sensitive results to prevent the reverse engineering of the original data. These mechanisms collectively ensure that computational results provide value without causing privacy breaches.

[0313] This step implemented the Medical Specialized Visualization System (MSVS), transforming complex analytical results into intuitive and easy-to-understand visual representations. MSVS is designed with a series of specialized visualization components tailored to the characteristics of medical data analysis: Clinical prediction visualization, displaying disease risk prediction and prognostic assessment, supporting risk stratification and time-series change analysis; Statistical association visualization, displaying multivariate correlations and causal relationships, supporting interactive exploration and hypothesis testing; Population distribution visualization, displaying patient population characteristic distribution and subgroup analysis, supporting multidimensional feature comparisons; and Treatment effect visualization, displaying the effectiveness evaluation and control analysis of interventions, supporting personalized treatment decisions. These specialized components enable medical professionals to quickly understand and apply analytical results.

[0314] To enhance the user experience, an Interactive Analysis Workbench (IAW) was designed, supporting exploratory analysis while protecting user privacy. IAW implements a "secure interactive analysis" paradigm, where user interactions are transformed into privacy-preserving computational tasks, and the results are processed securely before being displayed. Core functionalities include: hypothesis testing tools, allowing users to propose and verify medical hypotheses; feature importance analysis, showcasing key factors influencing prediction results; patient similarity search, finding similar cases while protecting privacy; and treatment comparison tools, evaluating the expected effects of different treatment options. This interactive design transforms the system from a simple computational tool into an intelligent assistant supporting medical decision-making.

[0315] To ensure system usability and user satisfaction, a User-Centered Design Approach (UCDA) was implemented, deeply integrating end-users into the design process. UCDA comprises four phases: user research to understand the workflows and needs of healthcare professionals; prototyping to create and iterate interactive prototypes; usability testing, inviting real-world users to evaluate the system and provide feedback; and continuous improvement, constantly optimizing the interface and functionality based on user feedback. Through this process, the system interface and functionality are closely integrated with medical practice, ensuring high usability and acceptability.

[0316] By integrating the aforementioned components and design methodologies, a medical data privacy computing fusion system was constructed. This system organically combines the aforementioned modules, forming a complete closed loop from data encryption, task parsing, multi-technology scheduling and execution to secure result display. The system realizes the privacy computing concept of "usable but invisible" medical data, enabling medical institutions to fully explore the value of data while protecting patient privacy, thereby promoting the development of medical research and clinical decision-making.

[0317] An application programming interface and visualization system were designed for medical researchers, providing a user-friendly interface and rich data analysis tools. It supports secure decryption and intuitive display of calculation results, forming a complete medical data privacy computing fusion system. This enables medical professionals without technical backgrounds to easily use privacy computing technology for secure medical data analysis.

[0318] In one embodiment of the present invention, step S11 involves analyzing the state space generated during the privacy computation process, determining the state characteristics and distribution patterns, and designing a prefix tree structure suitable for privacy computation of medical data, including node representation, path encoding, and tree balancing strategies, to obtain a prefix tree structure design scheme.

[0319] Specifically, in step S111, based on the data type, data scale, and access pattern in the multi-source medical data characteristics, the state characteristics and distribution patterns in the medical data state space are analyzed, the number of states, the frequency of state transitions, and the state dependencies are determined, and a medical data state space characteristic model is obtained.

[0320] First, a systematic analysis of the state space generated during medical data privacy computation is conducted to reveal its core characteristics and distribution patterns. State space refers to the set of all possible states a system may be in during computation, including intermediate computation results, temporary variables, and control information. In medical data privacy computation scenarios, the state space exhibits unique characteristics, and a deep understanding of these characteristics is crucial for designing efficient data structures.

[0321] First, by analyzing the multi-source medical data processing flow, three main state types were identified: data state, representing raw or transformed medical data, such as patient records, test results, or imaging data; computation state, representing intermediate results of algorithm execution, such as gradient values, partial aggregation results, or intermediate statistics; and control state, representing control information of the computation flow, such as task scheduling markers, synchronization points, or access permissions. These three states intertwine during privacy-preserving computation, forming a complete state space.

[0322] Regarding the characteristics of the number of states, analysis of execution logs for typical medical privacy computing tasks (such as multi-center joint modeling, security statistical analysis, and privacy-preserving queries) reveals a significant sparsity in the state space. Theoretically, the possible state space is extremely large (typically exponentially larger than the data size), but the states actually used occupy only a tiny fraction of the theoretical space, usually no more than 0.01%. This high sparsity means that traditional structures such as arrays or hash tables may result in severe space waste, while compressed structures such as prefix trees are more suitable.

[0323] To assess the state distribution characteristics, cluster analysis and entropy measurement methods were employed to reveal the locality of medical data states. State identifiers (typically multi-dimensional vectors containing information such as data source, processing stage, and operation type) exhibit a clear clustering phenomenon in space, with related states often sharing prefixes. This locality manifests as follows: different data states of the same patient tend to share identifier prefixes; intermediate result state identifiers from the same computational stage often have common parts; and control states of similar medical tasks often have similar identifier structures. This locality characteristic allows the prefix-sharing structure to effectively reduce storage redundancy.

[0324] To analyze the state transition patterns, a state transition graph was introduced, revealing a significant hierarchy and directionality in state transitions during healthcare data privacy computation. State transitions typically follow a predefined computational flow, gradually transforming from the input data state to intermediate computational states, ultimately reaching the result state. This hierarchy manifests in several ways: states can be naturally stratified, with clear dependencies between different levels; upper-level states often depend on combinations of multiple lower-level states; and state transitions primarily occur between adjacent levels, with cross-level transitions being relatively rare. This hierarchical structure closely aligns with the hierarchical organization of prefix trees.

[0325] The analysis also revealed state access patterns, showing significant temporal locality and access imbalance in privacy-preserving computations of medical data. Recently created states were accessed far more frequently than earlier states, and certain key states (such as aggregation results or global model parameters) were accessed frequently, while many intermediate states were accessed only once or infrequently. Specifically, approximately 20% of the states handled 80% of the access requests, consistent with a typical Pareto distribution. This access pattern suggests that prefix tree design should optimize hotspot paths and consider caching mechanisms to accelerate the retrieval of frequently accessed states.

[0326] By integrating the above analysis results, a characteristic model of the medical data state space was constructed, which is represented by a five-dimensional vector: <sparseness, locality measure, hierarchical depth, access distribution, and transition frequency>. The model quantitatively describes the core characteristics of the medical data privacy computing state space, providing a theoretical basis for the subsequent design of the prefix tree structure. Specifically, the model reveals three key characteristics of the medical privacy computing state: high sparsity, significant locality, and explicit hierarchy. These characteristics all point to prefix trees as an ideal data structure choice.

[0327] By analyzing the state space generated during medical data processing, it was found that the state of medical privacy computing has characteristics such as high sparsity (only a small portion of a large number of possible states are actually used), locality (related states are usually adjacent in the identifier space), and hierarchy (there are obvious hierarchical dependencies between states). These characteristics are very suitable for efficient storage using a prefix tree structure.

[0328] Step S112: Based on the medical data state space feature model, design a prefix tree node structure, including a node data field, a child node pointer array and a compression tag, design a path encoding scheme using variable length encoding, and design an AVL tree balancing strategy to maintain the query efficiency of the prefix tree, thus obtaining the prefix tree structure design scheme.

[0329] Based on the medical data state space features identified in the aforementioned steps, a trie structure specifically designed for privacy-preserving computations of medical data was developed. A trie is a tree-like data structure used for efficient string storage and retrieval; its core feature is reducing storage redundancy through shared prefixes. In this application scenario, the trie is extended and optimized to adapt to the specific needs of the medical data state space.

[0330] First, a prefix tree node structure adapted to the characteristics of medical data was designed. Traditional prefix tree nodes typically contain a fixed-size array of child node pointers, which is inefficient in the highly irregular state space of medical data, where branching factors vary greatly. To address this issue, an Adaptive Node Structure (ANS) was designed, comprising three core components:

[0331] The State Data Field (SDF) stores the actual state information associated with a node. Unlike traditional prefix trees, each node in a medical data prefix tree may contain actual data, not just leaf nodes. The SDF employs a layered design: a metadata layer stores descriptive information about the state, such as state type, creation time, and access counts; a data pointer layer stores pointers to the actual data blocks, supporting lazy loading; and a security tag layer stores privacy-related tags, such as access control lists and encryption flags. This design enables nodes to efficiently store different types and sizes of medical data states.

[0332] The Variable-length Children Array (VCA) is a core innovation of ANS, solving the problem of imbalanced branching factors in the medical state space. VCA replaces the fixed-size array of the traditional prefix tree, employing a dynamically resizable data structure that allocates space based on the actual number of child nodes. VCA includes three implementation modes, automatically selected by the system based on the number of child nodes: a simple array is used for a small number of child nodes (≤4); an ordered array with binary search is used for a medium number of child nodes (5-16); and a red-black tree is used for a large number of child nodes (>16). This multi-mode design maintains high efficiency under various branching conditions, avoiding wasted space and decreased search performance.

[0333] Path Compression Marker (PCM) is an optimization mechanism for handling long, single-branch chains in medical data state paths. In the medical data state space, long chains of paths with only a single child node often appear (such as a unique path for a specific patient ID). PCM allows these single-branch chains to be compressed into a single node, significantly reducing tree height and the number of nodes. PCM consists of two components: a compression flag, indicating whether the current node is a compression node; and a skip sequence, storing the compressed path segments. Through path compression, tree storage efficiency is improved by 35%-50%, while simultaneously accelerating path lookup.

[0334] Secondly, a variable-length path encoding scheme (VPE) for medical status paths was designed. Traditional prefix trees typically use fixed-length characters or numbers as path encoding, which is unsuitable for the diverse and hierarchical characteristics of medical data states. VPE employs a context-aware variable-length encoding strategy, dynamically adjusting the encoding length based on path hierarchy and semantics:

[0335] Level encoding divides the path into multiple semantic levels, each using a different encoding scheme. For example, medical institution identifiers use 2-byte encoding, data types use 1-byte encoding, patient IDs use variable-length encoding (4-12 bytes), and computation stages use 1-byte encoding. This hierarchical encoding reflects the natural organizational structure of medical data and improves encoding efficiency.

[0336] Entropy-optimized encoding allocates code length based on the frequency of occurrence of symbols at each level. High-frequency symbols (such as common data types or operations) are assigned shorter codes, while low-frequency symbols are assigned longer codes. This Huffman-like coding strategy reduces code length by an average of 25%, directly improving the storage efficiency of the prefix tree.

[0337] Incremental path compression uses the similarity between adjacent state paths for incremental encoding. Instead of storing the complete path, it stores the differences from the reference path. This incremental encoding is particularly suitable for time-related medical data states, such as consecutive test results for the same patient or iterative parameters of the same model.

[0338] Furthermore, an AVL Balancing Strategy (ABS) is introduced to maintain the performance stability of the prefix tree under frequent update operations. While traditional prefix trees do not require a balancing mechanism, dynamic state changes in healthcare data privacy computations can lead to severe imbalances in the tree structure, impacting query efficiency. ABS comprises three key mechanisms:

[0339] Local subtree balancing applies AVL balancing techniques to each subtree of the prefix tree. When the nodes in a subtree are unbalanced (some child node paths are significantly longer than others), the system reorganizes the subtree, ensuring that the longest path from any child node to a leaf node is no more than twice the length of the shortest path. This local balancing strategy avoids the high cost of full tree reorganization while maintaining query performance.

[0340] Hotspot path optimization adjusts the tree structure based on access frequency, shortening frequently accessed state paths. The system tracks the access count of each node and periodically reorganizes the tree structure, moving frequently accessed nodes closer to the root node. This optimization leverages the unevenness of medical data state access, reducing hotspot path query steps by an average of 35%.

[0341] Incremental rebalancing distributes the balancing operation across multiple small adjustments, avoiding the impact of a single large-scale reorganization on system performance. The system performs incremental rebalancing during idle periods, adjusting only a small portion of the tree each time, ensuring that the balancing operation does not affect the response time of normal computing tasks.

[0342] Finally, memory optimization techniques were integrated to further improve the storage efficiency of the prefix tree. Key optimizations include: node pool management, using pre-allocated node pools to reduce memory fragmentation and allocation overhead; a state sharing mechanism, allowing multiple paths to reference the same state data, avoiding data redundancy; and a tiered storage strategy, allocating state data to storage layers of different speeds based on access frequency, storing hot data in a high-speed cache and moving cold data to slower storage.

[0343] To address the state space characteristics of medical data, a dedicated prefix tree structure was designed: nodes contain data fields to store actual state information, variable-length child node pointer arrays support different branching factors, and compressed tags are used for path compression; a variable-length encoding scheme is adopted to efficiently represent the path information of state identifiers; an AVL balancing strategy is introduced to ensure the balance of the tree and query efficiency under a large number of state insertion and deletion operations, thus forming a prefix tree structure design scheme suitable for privacy computing of medical data.

[0344] In one embodiment of the present invention, step S23 involves modeling the privacy computing technology selection problem as a non-uniform k-center problem based on the medical computing task category model, defining a cost function and resource constraints, and obtaining a mathematical optimization model.

[0345] Specifically, in step S231, based on the medical computing task category model, all classified medical data computing tasks are extracted to form a medical data computing task set. Each task in the medical data computing task set is defined as a feature vector containing data scale, computational complexity, and privacy requirements. Privacy computing technologies, including federated learning, homomorphic encryption, and secure multi-party computation, are defined as a technology set to obtain a task-technology mapping framework.

[0346] Based on a medical computing task category model, a mathematical foundation for the selection of privacy-preserving computing technologies was constructed, formalizing the problem into a task-technology mapping framework. This framework abstracts medical computing tasks and privacy-preserving computing technologies into quantifiable and comparable mathematical entities, laying the foundation for subsequent optimization modeling.

[0347] First, this step systematically extracts features from medical data computation tasks, constructing a Medical Computation Task Feature Vector (MCTFV). MCTFV is a multi-dimensional feature representation of medical computation tasks, containing three main categories of feature dimensions:

[0348] The data feature dimension capture task involves the key characteristics of the data, including: data size (in records or bytes), ranging from small datasets (<1GB) to very large datasets (>1TB); data dimension, representing the number of features, ranging from low dimension (<10) to very high dimension (>10000); data distribution, describing the distribution pattern of data among multiple participants, including horizontal partitioning (samples distributed across institutions), vertical partitioning (features distributed across institutions), and mixed partitioning; data sparsity, representing the proportion of effective data, ranging from dense (>50%) to highly sparse (<0.1%); and data update frequency, ranging from static (no updates) to high-frequency dynamic (updated multiple times per second).

[0349] The computational feature dimension describes the algorithm and computational requirements of the task, including: computational complexity, representing the asymptotic time complexity of the algorithm, from linear O(n) to exponential O(2^n). n Parallelism potential, quantifying the degree to which a task can be parallelized, from low (<20% parallelizable) to high (>80% parallelizable); Memory intensity, representing the ratio of memory requirements to data size, from low (<2 times) to high (>10 times); Precision requirements, describing the required precision of the computation results, from low (allowing approximation) to high (requiring precision); Interaction rounds, representing the number of communication rounds required during the computation process, from single round to high-frequency multi-round.

[0350] Privacy features represent the privacy protection requirements of a task, including: Sensitivity level, indicating the degree of data sensitivity, ranging from low (anonymous statistical data) to extremely high (personal genetic information); Attack model, referring to the assumed threat model in the computing environment, ranging from semi-honest (following the protocol but attempting to infer information) to malicious (potentially deviating arbitrarily from the protocol); Information disclosure tolerance, indicating the acceptable level of information disclosure, ranging from high (allowing partial disclosure of statistical characteristics) to zero (not allowing any disclosure); Validation requirements, referring to the need to verify the correctness of the results, ranging from none (trusting the computation results) to strict (requiring zero-knowledge proofs); and Compliance requirements, indicating applicable regulatory standards, such as HIPAA, GDPR, etc.

[0351] Based on the multidimensional feature description described above, each medical computation task is represented as a feature vector, where each dimension d j The vectors are normalized to the [0,1] interval for uniform computation. For example, a multi-center joint prediction model training task might be represented as <0.7, 0.5, 0.8, 0.3, 0.1, 0.6, 0.7, 0.4, 0.8, 0.6, 0.9, 0.8, 0.3, 0.7,0.9>. This vectorized representation allows the task to be precisely located and compared in the feature space.

[0352] Secondly, this step systematically defines the Privacy-preserving Computation Technology Set (PCTS). The PCTS includes various privacy-preserving computation technologies and their variants, each represented by its characteristic vector:

[0353] The federated learning technology family includes: horizontal federated learning, suitable for horizontal data segmentation scenarios; vertical federated learning, suitable for vertical feature segmentation scenarios; federated transfer learning, a variant of federated learning that combines knowledge transfer; distributed federated learning, employing a fully decentralized architecture; and federated reinforcement learning, a variant of federated learning that integrates reinforcement learning. For each federated learning variant, further subdivisions are made into technical instances with different parameter configurations, such as combinations of different aggregation algorithms (FedAvg, FedProx, etc.) and different privacy enhancement mechanisms (differential privacy, secure aggregation, etc.).

[0354] Homomorphic encryption techniques include: partially homomorphic encryption, such as Paillier encryption (supporting addition) and ElGamal encryption (supporting multiplication); fully homomorphic encryption, such as the BFV scheme (suitable for integer computation) and the CKKS scheme (suitable for approximate real number computation); algebraic homomorphic encryption, supporting operations on specific algebraic structures; and ring homomorphic encryption, based on ring structures. Each homomorphic encryption technique is further subdivided into variants with different parameter settings, such as different security levels, different polynomial modulus sizes, and different optimization strategies.

[0355] Secure multi-party computation techniques include: circuit-based secure computation, such as Garbled Circuits and Boolean sharing; secret-sharing-based secure computation, such as Shamir secret sharing and additive secret sharing; oblivious transfer-based protocols; zero-knowledge proof techniques; and secure hardware-assisted multi-party computation, such as schemes based on Trusted Execution Environments (TEEs). These techniques are further subdivided into various technical instances based on different parameter settings and optimization strategies.

[0356] Each technology is represented as a technology characteristic vector, capturing its key characteristics, including dimensions such as computational efficiency, communication overhead, memory requirements, security strength, flexibility, and implementation maturity. For example, fully homomorphic encryption technology based on CKKS might be represented as <0.2, 0.3, 0.2, 0.9, 0.7, 0.8, 0.9, 0.4, 0.6, 0.5, 0.7, 0.8>, indicating its characteristics of low computational efficiency but high security, and support for complex operations but high communication overhead.

[0357] This step innovatively constructs the Task-Technology Mapping Framework (TTMF), the core structure connecting medical computing tasks and privacy-preserving computing technologies. TTMF comprises three key components:

[0358] The feature space definition unifies the task feature vector space and the technical characteristic vector space into a multi-dimensional mapping space, making the two comparable. This involves feature dimensionality reduction and spatial transformation to ensure that the two types of vectors are comparable in the transformed space.

[0359] Compatibility function, defining task t i Technology h j The compatibility measure between them is comp(t) i ,h j The compatibility function reflects the suitability of technology for the task. It considers functional matching (whether the technology can achieve the functions required for the task) and performance matching (whether the performance of the technology meets the requirements of the task). The output value ranges from [0,1], and the higher the value, the more compatible the technology is.

[0360] The mapping rule set defines the set of rules that specify which technology a task should choose under certain conditions. These rules are based on expert knowledge and historical experience, and take the form of "IF condition THEN recommended technology", such as "IF task involves high-dimensional data AND high sensitivity level AND attack model is semi-honest THEN recommended horizontal federated learning + differential privacy".

[0361] The TTMF framework provides a conceptual foundation and initial inspiration for subsequent optimization modeling by mapping abstract task characteristics and technical features to specific selection recommendations. This framework specifically considers the unique needs of medical data processing, such as stringent privacy requirements, high-dimensional sparse data characteristics, and complex computational patterns, ensuring that technology selection meets the specific challenges of medical scenarios.

[0362] Medical computing tasks are abstracted into feature vectors, including dimensions such as data scale (e.g., data volume), computational complexity (e.g., algorithm complexity), and privacy requirements (e.g., sensitivity level). Available privacy computing technologies (including variants with different parameter configurations) are defined as a set of technologies, and a framework for mapping relationships between tasks and technologies is established.

[0363] Step S232: Based on the task-technology mapping framework, define a distance function to represent the cost of executing tasks in the medical data calculation task set using technologies in the technology set, define a capacity constraint for each technology in the technology set to represent the maximum number of tasks that each technology can process simultaneously, and a radius constraint to represent the maximum acceptable execution cost of each technology, thus obtaining a set of constraints.

[0364] Building upon the task-technology mapping framework constructed in the preceding steps, this paper further defines a quantified distance function and constraints, laying the foundation for mathematical optimization modeling of the privacy computing technology selection problem. This step transforms the abstract mapping relationship into a computable mathematical model, enabling the optimization problem to be precisely expressed and solved.

[0365] First, a Medical Task-Technology Distance Function (MTTDF) was designed. This is a multidimensional weighted distance metric that represents the cost incurred when a medical computing task is performed using specific privacy-preserving computing technologies. MTTDF considers various cost factors and assigns different weights based on the characteristics of the medical scenario, forming a comprehensive distance metric.

[0366] Computation time cost reflects the processing time required to execute a task. Factors to consider include: original computational complexity, the complexity of the task algorithm itself; additional overhead introduced by technology, such as additional computation introduced by encryption operations and security protocols; the impact of data size, the influence coefficient of data volume on computation time; and the potential for parallel acceleration, the degree of support of technology for parallel computing.

[0367] Communication overhead cost represents the amount of data transmitted and the number of communication rounds during task execution. Factors to consider include: number of protocol interaction rounds, the number of communication rounds required to complete the calculation; amount of data transmitted, the total amount of data that needs to be transmitted between the participants; bandwidth utilization, the degree to which the technology effectively utilizes available bandwidth; and latency sensitivity, the degree to which the calculation process is sensitive to network latency.

[0368] The quantification of memory and storage costs considers factors such as: peak memory usage (maximum memory consumption during computation); intermediate result storage requirements (size of intermediate computation results to be saved); memory access patterns (different efficiencies due to sequential or random access); and cache efficiency (efficiency of technology in utilizing processor cache).

[0369] Security costs reflect the suitability of technology in meeting mission security requirements. Factors to consider include: security strength gap (the gap between the security level provided by the technology and the mission requirements); privacy risk (the potential extent of information leakage); attack model coverage (whether the technology can cope with the attack models specified for the mission); and security proof level (the strength of the formal proof of the technology's security).

[0370] Flexibility and adaptability cost refers to the ability of a technology to adapt to the specific needs of a task. Factors to consider include: functional coverage, the degree to which the set of operations supported by the technology matches the task requirements; accuracy guarantee, whether the technology can meet the computational accuracy requirements of the task; scalability, performance as the data scales up; and fault tolerance, the ability to cope with participant failures or network problems.

[0371] Based on the aforementioned multidimensional cost factors, MTTDF is defined as a weighted cosine distance, and the task feature vector t is calculated. i Technical characteristic vector h j The distance between them is denoted as d(t). i ,h j Weighted cosine distance takes into account the angular differences and directional consistency between vectors, making it more suitable than Euclidean distance for capturing the degree of matching between tasks and technologies. The weights of each dimension in the distance calculation are determined by analyzing historical task execution data and expert evaluations. For example, in privacy analysis of genomic data, the security dimension is given higher weight; while in real-time medical monitoring analysis, the time dimension is given higher weight.

[0372] To address the unique challenges of healthcare scenarios, MTTDF also incorporates adaptive adjustment mechanisms. These mechanisms adjust the base distance calculation using contextual factors, taking into account situational factors such as urgency (emergency analysis vs. routine research), resource fluctuations (peak hours vs. off-peak hours), and changes in regulatory requirements. For example, for healthcare analytics tasks marked as "urgent," the weight of the time dimension is automatically increased; conversely, in the face of stringent new data protection regulations, the weight of the security dimension is increased accordingly.

[0373] Secondly, this step defines two types of key constraints for each privacy computing technology, forming a Technology Constraint Set (TCS):

[0374] A capacity constraint represents the maximum number of tasks a specific technology can handle simultaneously, reflecting the upper limit of its processing power. Capacity constraints consider multiple factors: computational resource limitations, such as available CPU cores, GPU units, and memory capacity; concurrent processing capability, the maximum number of parallel tasks supported by the technology; licensing restrictions, such as concurrent usage license limitations for certain commercial technologies; and system stability, the maximum load required to maintain stable system operation. A capacity constraint is defined as the maximum number of tasks cap(h) for technology j. j For example, a fully homomorphic encryption system with a specific configuration may be limited to handling a maximum of five computationally intensive tasks at the same time to ensure performance and stability.

[0375] A radius constraint represents the maximum acceptable task-technology distance for a technology, defining the limits of its applicability. The radius constraint reflects a key concept: when the "distance" between the task and the technology (i.e., the cost of using the technology to perform the task) exceeds a certain threshold, the technology is no longer a feasible option for that task. The radius constraint considers: a performance degradation threshold (the distance at which the technology's performance begins to significantly decline); a resource consumption ceiling (the maximum acceptable resource consumption); a safety assurance floor (the minimum safety standard that the technology can guarantee); and a quality assurance limit (the maximum distance at which the quality of the results can be maintained). The radius constraint is defined as the maximum acceptable distance (rad(h)) for technology j. j For example, a specific federated learning scheme might set rad(h) j A value of 0.7 indicates that when the task-technology distance exceeds 0.7, the technology is no longer applicable, possibly because security cannot be guaranteed or performance is severely degraded.

[0376] In addition to technology-specific constraints, this step also defines system-level global constraints (SGC) to capture the limitations of the entire medical data processing system:

[0377] Overall resource constraints limit the total computing, storage, and network resources available to the system, ensuring that the sum of resource usage across all technologies does not exceed the system capacity. This includes: CPU / GPU total computing power limits; total memory and storage space limits; network bandwidth limits; and energy consumption limits (especially for mobile medical devices).

[0378] Compliance constraints ensure that technology choices meet applicable regulatory and standard requirements, including: data residency requirements, where data must be stored in a specific geographic location; processing transparency requirements, where processing must be auditable; algorithm interpretability requirements, where decision-making processes must be understandable and explainable; and disaster recovery requirements, where the system must be able to recover from failures.

[0379] Load balancing constraints promote the balanced use of system resources and prevent overloading of certain technologies or infrastructure components. These include: cross-technology load distribution requirements; cross-hardware node balancing requirements; peak load management strategies; and quality of service (QoS) assurance requirements.

[0380] By integrating MTTDF and TCS, this step constructs a complete set of constraints, providing the necessary elements for mathematical modeling of the optimization problem. These constraints not only consider the direct costs and limitations of technology selection but also incorporate requirements specific to the healthcare environment, such as data protection regulations, timeliness of medical decision-making, and system reliability, ensuring that the final technology selection optimizes performance while meeting the specific needs of the healthcare scenario.

[0381] A multidimensional distance function is defined to quantify the "cost" of a task when executed using a specific technology, including a weighted combination of factors such as computation time, memory usage, and communication overhead. Two types of constraints are defined for each technology: a capacity constraint, which represents the maximum number of tasks that the technology can handle simultaneously, and a radius constraint, which represents the maximum task distance (cost) that the technology can accept. These constraints reflect the processing capacity and applicability limitations of the technology.

[0382] Step S233: Based on the set of constraints, construct the objective function as minimizing the largest task-technology distance among all tasks, and set the constraints as follows: the number of tasks assigned to each technology does not exceed the corresponding capacity constraint and the distance between each task and the corresponding assigned technology does not exceed the radius constraint of the corresponding technology, thus obtaining the mathematical optimization model.

[0383] Based on the distance function and constraint set defined in the preceding steps, the privacy computing technology selection problem is formally modeled as a non-uniform k-center problem, a variant of a classic combinatorial optimization problem. This modeling approach transforms the abstract technology selection into a structured mathematical optimization problem, making it solvable through optimization algorithms.

[0384] First, the correspondence between the non-uniform k-center problem and the privacy computing technology selection problem is clarified. The traditional k-center problem seeks to assign a set of points to k centers such that the maximum point-center distance is minimized. The extension of the non-uniform k-center problem lies in the fact that each center has different capacity and service radius constraints. In the context of medical privacy computing: medical data computing tasks correspond to the point set, with each task being a point in the feature space; privacy computing technologies correspond to the centers, with each technology serving multiple tasks; the task-technology distance corresponds to the point-center distance, representing the cost of executing a task using a specific technology; the technology capacity constraint corresponds to the center capacity limit, representing the maximum number of tasks that each technology can serve simultaneously; and the technology radius constraint corresponds to the center service radius, representing the maximum acceptable service distance for the technology. This correspondence allows the medical privacy computing technology selection problem to be formally expressed and solved using the theoretical framework of the non-uniform k-center problem.

[0385] Secondly, this step constructs a mathematical optimization model that precisely expresses the objective function and constraints of the problem:

[0386] The objective function is set to minimize the maximum task-technology distance among all tasks, i.e., the minimax optimization objective: min max{d(t i ,h j ) | t i Assigned to h j This objective function reflects a "bottleneck optimization" strategy, focusing on the task-technology pairing with the highest execution cost in the system and attempting to reduce this maximum cost to improve the overall system's efficiency floor. This strategy is particularly suitable for medical scenarios because a single extremely slow task can delay the entire medical analysis process, impacting clinical decision-making.

[0387] The constraint set contains four key types of constraints: allocation integrity constraints, which require that each task must be assigned to one and only one technology to ensure that all tasks can be processed; capacity constraints, which require that the total number of tasks assigned to each technology does not exceed the capacity limit cap(h) of that technology. j To prevent technology overload; radius constraints require that the distance between a task and its assigned technology does not exceed the upper limit of the technology's radius in rad (h). j This ensures that the technology is suitable for the assigned task; global resource constraints require that the total resources used by all technologies do not exceed the upper limit of the system's available resources, thus ensuring system feasibility.

[0388] By introducing decision variable x ijThe complete mathematical optimization model is formalized using the variable z (representing whether task i is assigned to technology j) and the auxiliary variable z (representing the maximum task-technology distance). This model is a mixed-integer programming problem, and its complexity stems from the duality of the decision variables (tasks are either assigned to a specific technology or not) and the minimax property of the objective function.

[0389] To address the computational complexity of the model, this step proposes a problem characteristic analysis and algorithm selection strategy:

[0390] Complexity analysis shows that the non-uniform k-center problem is NP-hard, meaning that finding the exact global optimum is often computationally infeasible for large-scale instances. In healthcare privacy computing scenarios, the number of tasks can reach hundreds to thousands, with dozens of technical variations, further expanding the problem scale.

[0391] Problem structure analysis reveals that the selection of medical privacy computing technologies has several advantageous characteristics: sparse connectivity, meaning that not all tasks can be handled by all technologies, and many task-technology pairs are automatically excluded due to radius constraints; technology hierarchy, meaning that there are inclusion relationships between technologies, with higher-level technologies often covering the functions of lower-level technologies; and local independence, meaning that some task groups are independent of each other and can be decomposed into subproblems that can be solved in parallel.

[0392] Based on these characteristics, an algorithm selection framework is proposed, which selects appropriate solution methods according to the problem size and structural features: for small-scale problems (<50 tasks, <10 techniques), exact algorithms such as branch and bound can be used; for medium-scale problems, approximate algorithms such as greedy algorithms plus local search are recommended; for large-scale problems, metaheuristic algorithms such as simulated annealing or genetic algorithms, or problem decomposition methods to break down large problems into manageable subproblems are recommended.

[0393] Furthermore, the dynamic characteristics of the healthcare environment were considered, and the model was extended to adapt to changing needs:

[0394] Dynamic task processing extensions allow the model to adapt to new tasks or changes in task priorities. When a new medical analysis task is submitted, the system can incrementally update the assignment scheme instead of resolving the entire problem. This dynamic adaptability is particularly important for responding to urgent medical analysis needs.

[0395] Elastic technology capacity scaling takes into account the possibility that technology resources may change over time. For example, in a cloud environment where computing resources are scalable, the capacity constraints of a technology may be elastic, allowing scaling during periods of high demand. The model captures this elasticity by setting the capacity parameter as a function that varies over time, rather than a fixed constant.

[0396] Robustness considerations extend the model's ability to handle parameter uncertainties. Task characteristics and technical performance in medical settings may be subject to measurement errors or fluctuations. By incorporating scenario analysis and worst-case optimization, the model ensures that the selected approach remains feasible and efficient under various possible conditions.

[0397] Finally, the applicability of the mathematical model was verified through evaluation using real-world medical data analysis tasks and various privacy-preserving computing technologies. The verification results show that the model effectively captures the core needs and constraints of medical privacy-preserving computing scenarios, and the generated technology allocation scheme improves overall performance by an average of 25%-40% compared to manual selection or simple heuristic methods, while ensuring that all security and resource constraints are met.

[0398] Through mathematical modeling in this step, the problem of choosing privacy-preserving computation technologies is transformed from an abstract decision-making process into a structured optimization problem, which can be systematically solved using mature optimization theories and algorithms. This transformation makes the complex technology selection process more transparent, reproducible, and optimizable, providing a solid theoretical foundation for improving the efficiency and security of privacy-preserving computation in medical data.

[0399] Based on the defined task-technology distance and constraints, the optimization objective is set as minimizing the maximum execution cost (maximum task-technology distance) among all tasks, while satisfying the constraints of technology capacity and applicability. A complete non-uniform k-center mathematical optimization model is constructed, and the privacy computing technology selection problem is formalized as the process of solving this model.

[0400] In one embodiment of the present invention, step S34 involves designing and implementing a minimum weight basis solution algorithm based on the element weight matrix, selecting a combination of protocol elements from the protocol elements whose differential privacy parameter ε value is less than a preset security threshold and whose total computational cost and communication overhead are minimized, to obtain the protocol minimum weight basis.

[0401] Specifically, in step S341, based on the computational cost, communication overhead, and security strength weight of each protocol element defined in the element weight matrix, all protocol elements in the element weight matrix are extracted to create a protocol element selection set, and all protocol elements are sorted in descending order of security strength to obtain a candidate element sequence.

[0402] First, a set of protocol elements is extracted from the constructed element weight matrix and prioritized based on security metrics. The element weight matrix is ​​a multi-dimensional data structure describing the attributes of each component of the privacy computing protocol, containing key metrics such as computational cost, communication overhead, and security strength for each protocol element.

[0403] First, all available protocol elements are extracted from the element weight matrix to form an initial candidate set. Protocol elements are the fundamental components of a privacy-preserving computation protocol, including: cryptographic primitives such as homomorphic encryption, blind signatures, or zero-knowledge proofs; communication mechanisms such as secure channels or obfuscated circuits; computational components such as secure aggregation or distributed query processing; and privacy enhancement techniques such as differential privacy, k-anonymity, or randomized responses. For privacy-preserving computation of medical data, a typical set of elements contains 30-50 different protocol elements, each with specific functions and performance characteristics.

[0404] Next, the security strength index for each element is extracted, primarily based on the differential privacy parameter ε. The differential privacy parameter ε is a key indicator for quantifying the strength of privacy protection; a smaller ε value indicates a higher level of privacy protection, but it usually also implies higher computational costs or lower data availability. For example, ε=0.1 represents extremely strong privacy protection, while ε=10 represents relatively weak protection. For each protocol element, the system evaluates its native ε value or its contribution to the overall protocol ε value.

[0405] In addition to the ε value, a comprehensive security strength score is constructed by considering multiple security indicators, including: resistance model strength (the types of attacker models the element can resist, such as semi-honest, malicious, or side-channel attacks); computational security (the strength of cryptographic assumptions and the strictness of security proofs); communication security (the degree of protection of communication channels and resistance to eavesdropping); and information leakage (the maximum information leakage the element may cause). These indicators are weighted and combined to form the comprehensive security strength score for each element.

[0406] A security priority ranking algorithm was implemented to sort all protocol elements from highest to lowest security strength. The ranking process employs a two-stage strategy: First, a primary ranking is performed based on the ε value, with elements having smaller ε values ​​ranking higher; second, for elements with the same or similar ε values, a secondary ranking is performed based on a comprehensive security strength score. The ranking algorithm considers the trade-offs between security metrics. For example, some elements may have extremely low ε values ​​but pose a risk of side-channel attacks; in such cases, the system will adjust the final ranking according to the importance of the application scenario.

[0407] The sorting process generates a sequence of candidate elements, starting with the most secure elements and gradually transitioning to elements with lower security but potentially higher computational efficiency. The front of the sequence typically contains elements with strong differential privacy guarantees, such as those using low-ε randomization response mechanisms or noise-addition techniques; while the back of the sequence contains elements with higher computational efficiency but relatively weaker security guarantees. This sorting ensures that security requirements are prioritized in subsequent selection processes.

[0408] Extract all protocol elements from the element weight matrix to create an initial selection set; sort the elements from high to low according to the security strength index (mainly the differential privacy parameter ε value) to form a priority-ordered candidate element sequence, ensuring that security requirements are given priority.

[0409] Step S342: Based on the candidate element sequence, a greedy strategy is used to sequentially select protocol elements that satisfy the differential privacy parameter ε value being less than a preset security threshold, calculate the total computational cost and communication overhead of the currently selected element combination, and obtain the initial protocol element combination.

[0410] Based on the candidate element sequence generated by the preceding steps, an initial protocol element combination is constructed using a greedy strategy. The greedy strategy is an iterative decision-making method that makes what appears to be the optimal choice at each step, aiming to ultimately obtain a globally optimal or near-optimal solution.

[0411] First, a differential privacy security threshold εmax is set, which is the maximum ε value allowed for protocol elements. The security threshold is set based on the privacy requirements of the application scenario and relevant regulatory requirements. For example, when processing highly sensitive medical data (such as genetic information or mental health records), εmax may be set to 0.5 or lower; while when processing less sensitive aggregated statistical data, εmax may be relaxed to 2 or 3. The threshold setting also considers the data scale and query complexity. Simple queries on larger datasets can use lower thresholds, while complex analyses or small datasets may require appropriate relaxation.

[0412] Next, initialize an empty set of element combinations S, and define the necessary set of functional requirements F, including all the functions that the protocol must support, such as secure aggregation, conditional querying, and model training. Each functional requirement may be satisfied by multiple different protocol elements, but at least one element must implement it.

[0413] The greedy selection process is performed iteratively. In each iteration, element e is considered sequentially from the candidate element sequence, and it is checked whether it meets three conditions: security condition, the differential privacy parameter ε(e) of e ≤ εmax, ensuring that the element meets the security threshold requirement; functional condition, e can satisfy at least one of the currently uncovered functional requirements; and compatibility condition, e is compatible with all elements in the selected element set S, and there are no functional or security conflicts.

[0414] For an element that satisfies these three conditions, the marginal cost increase after adding the element is calculated, including the computational cost increment ΔComp and the communication overhead increment ΔComm. The calculation considers the synergistic effects between elements; certain element combinations may produce additional optimizations or additional overhead. The marginal cost is calculated through a weighted combination: ΔCost = w1·ΔComp + w2·ΔComm, where w1 and w2 are weights reflecting the relative importance of computational and network resources.

[0415] Select the element that covers the unmet functional requirements with the lowest marginal cost and add it to S. If multiple elements have the same marginal cost, prioritize the element with higher safety strength. After each element is added, update the set of uncovered functional requirements and the current total cost.

[0416] The greedy process continues until all functional requirements are met or no new element can be found that satisfies the conditions. In the latter case, the security threshold or functional requirements will be relaxed, or the user will be prompted that the protocol cannot be constructed under the current constraints.

[0417] After the greedy selection is completed, an integrity check is performed on the initial combination of elements to ensure that all necessary functions are covered. The check includes function coverage verification to ensure that each functional requirement is supported by at least one element; security compatibility verification to ensure that there are no security conflicts between all selected elements; and resource requirement assessment to calculate the total computational cost and communication overhead of the combination to ensure that it is within the system resource capacity.

[0418] The key innovation in this step is a function-driven, security-first selection strategy. This means that the most secure elements are considered first, and only when these elements cannot meet a functional requirement are elements with lower security but still meeting the required threshold considered. This ensures that the protocol uses the most secure combination of elements possible while still fulfilling functional requirements.

[0419] Using a greedy strategy, an initial combination of protocol elements was constructed that satisfies all functional requirements and security thresholds while minimizing total computational cost and communication overhead. This initial combination will serve as the starting point for subsequent optimizations.

[0420] A greedy strategy is adopted to sequentially select elements from the sorted candidate element sequence whose differential privacy parameter ε values ​​meet the security threshold requirements. The impact of adding each element to the current combination on the total cost is evaluated, and elements that can meet the functional requirements and increase the minimum total cost are selected until a complete initial protocol element combination is formed.

[0421] Step S343: Based on the initial protocol element combination, perform element replacement and combination optimization through an iterative approach. While ensuring that the differential privacy parameter ε value meets the security threshold, continuously reduce the total computational cost and communication overhead. Stop the iteration when the total computational cost and communication overhead cannot be reduced for several consecutive iterations, and obtain the locally optimal element combination.

[0422] Based on the initial protocol element combination constructed in the previous steps, implement an iterative optimization process to further reduce costs. Iterative optimization is a method to improve the initial solution by continuously making small adjustments to the current solution and gradually approaching the optimal solution.

[0423] First, define the optimization objective function to quantify the total cost of the element combination: Cost(S) = α·CompCost(S) + β·CommCost(S) + γ·OverheadCost(S), where CompCost is the computational cost, CommCost is the communication overhead, OverheadCost is the management overhead, and α, β, γ are weight factors reflecting the relative importance of different resource dimensions. In the medical data processing scenario, the computational cost is usually measured by processor time and memory usage; the communication overhead is measured by the data transfer volume and the number of communication rounds; the management overhead includes additional costs such as key management and synchronization coordination.

[0424] Adopt a local search optimization strategy to explore the "neighborhood" of the current solution to find a better solution. The neighborhood is defined as all possible combinations that can be reached through single element replacement, element pair exchange, or adjustment of a finite set of elements. Specifically, the system implements three types of local improvement operations:

[0425] The element replacement operation considers replacing a certain element in the current combination with other elements in the candidate set, provided that the function requirements and security threshold are still met after the replacement. Enumerate each element e1 in the current combination and try to replace it with each element e2 within the security threshold in the candidate set. If e2 can cover all the functions of e1 and Cost(S - {e1} + {e2}) < Cost(S), then perform the replacement. For example, the system may replace an exact but computationally intensive differential privacy mechanism with a computationally more efficient approximate differential privacy mechanism, provided that the security parameters still meet the threshold.

[0426] The element combination optimization considers replacing multiple elements in the current combination with an alternative group that is functionally equivalent but has a lower total cost. Identify subsets of elements with overlapping functions and explore whether there are alternative solutions that can provide the same functions at a lower cost. For example, it may replace separate encryption, aggregation, and decryption elements with a comprehensive secure aggregation element, reducing the coordination overhead and communication cost.

[0427] Redundancy elimination analyzes functional redundancy in the current composition and removes unnecessary elements. It checks whether the functionality of each element is completely covered by other elements; if so, and removal does not affect overall security, it is discarded. This optimization is particularly suitable for redundancy that may be introduced during a greedy construction process.

[0428] In each iteration, all possible local improvement operations are evaluated, and the operation that produces the greatest cost reduction is selected for execution. If multiple operations produce the same cost reduction, the operation with the greatest safety improvement is prioritized. After each operation, the system updates the current combination of elements and the total cost.

[0429] A dynamic security constraint check mechanism was introduced during the optimization process to ensure that any element change does not violate security requirements. This mechanism calculates the effective differential privacy parameter εeff before and after the combination change, ensuring that εeff ≤ εmax. The εeff calculation takes into account the combination effect between elements; some element combinations may produce cumulative security effects or security weakening.

[0430] To avoid local optima traps, a variant of the simulated annealing strategy is implemented. In some iterations, there is a small probability of accepting an operation that slightly increases cost but may lead to a better solution. The probability of acceptance decreases with the increase in cost and the current iteration number, ensuring that the algorithm eventually converges.

[0431] The iterative process has a termination condition: when the total cost cannot be reduced by more than a predetermined threshold δ (usually set to 0.5%) after k consecutive iterations (typically k=5 to 10), a local optimum is considered reached. Furthermore, a maximum number of iterations, max, is set. iter This serves as a backup termination condition to prevent the algorithm from getting stuck in a loop.

[0432] After iterative optimization, a locally optimal combination of elements is output. This combination satisfies the safety threshold requirement and significantly reduces the total cost of the initial combination through multiple rounds of optimization. Experiments show that iterative optimization can typically reduce the cost of the initial combination by 15%-30% while maintaining or improving safety.

[0433] Based on the initial combination of elements, the solution is continuously improved through an iterative optimization process: trying to replace existing elements with lower-cost alternatives, exploring possible adjustments to the combination of elements, evaluating the impact of replacements and adjustments on the total cost in each iteration, accepting changes that can reduce costs while ensuring that the safety threshold is not violated, and considering that a local optimum has been reached when multiple iterations fail to further reduce costs.

[0434] Step S344: Based on the locally optimal combination of elements, verify the independence constraints and matroid properties of the selected protocol elements to ensure that the combination of elements meets the requirements of protocol integrity and security, and obtain the minimum weight basis of the protocol.

[0435] Based on the locally optimal element combinations obtained from the preceding steps, a comprehensive verification is performed to confirm that they satisfy all requirements of the protocol's minimum weight basis. Minimum weight basis verification is a crucial step in ensuring that the optimization results meet theoretical guarantees, involving independence constraint checks and matroid property verification.

[0436] First, the independence constraints of the selected element combination are verified. Independence constraints mean that there is no functional redundancy or circular dependency in the element set, and each element has an irreplaceable role. The system implements three levels of independence verification:

[0437] The functional independence check ensures that each element provides at least one function that cannot be completely substituted by any other element. A function coverage matrix M is constructed, where M[i,j]=1 indicates that element i supports function j. Then, it is verified that removing any element results in at least one function being lost. This check prevents the existence of redundant elements and ensures the minimization of combinations.

[0438] Security independence analysis ensures that each element makes a unique contribution to overall security. The impact of removing each element on the combined differential privacy parameter εeff is evaluated; if removal leads to a significant increase in εeff (a decrease in security), then that element possesses security independence. Some elements may not provide differential privacy protection on their own, but may synergistically enhance overall security with other elements.

[0439] Resource independence assessment ensures that there is no unnecessary duplication in resource utilization among the element combinations. It examines the possibility of resource sharing and reuse among elements to ensure that there are no duplicate elements that can be merged to reduce resource consumption.

[0440] Next, we verify the matroid properties of the element combinations, which is the theoretical foundation for ensuring that the combinations form an efficient minimum weight basis. A matroid is a structure in combinatorics with specific properties that make it suitable for describing the independence relationships between elements. In the context of privacy-preserving computation protocols, the matroid model views the protocol as a set of elements and independence constraints. We systematically verify three key matroid properties:

[0441] The heritability check ensures that subsets of any independent element subset are also independent. Verify any subset S' in the combinatorial set. Does S also satisfy the independence constraint, meaning that no element in a subset can be completely replaced by any other element in that subset? This property ensures that there are no hidden redundant levels in the combination.

[0442] The commutativity property verification ensures that the combination of elements possesses optimal commutativity. For any independent set I and any independent set J larger than I, there is always an element from J that can be added to I to maintain independence. The system verifies this property by trying various element swap operations, ensuring that the combination cannot be further optimized by simple swaps.

[0443] Maximality verification ensures that the combination of elements is a maximal set in terms of inclusion relations, meaning that no more elements can be added while maintaining independence and satisfying the safety threshold. Try adding each remaining element from the candidate set to the current combination and verify whether it would cause a violation of the independence constraint or exceed the safety threshold.

[0444] After verifying the matroid properties, the minimum weight basis is finally confirmed. The minimum weight basis is the basis with the smallest total weight (cost) among all satisfying bases. The total weight W(S) of the current combination is calculated and compared with the theoretical lower bound. Although precisely calculating the optimal solution is usually computationally too complex, the system uses approximate comparisons and lower bound estimates to evaluate the quality of the current solution and confirm that it is close to the global optimum.

[0445] Finally, a comprehensive verification of the protocol's integrity and security is conducted, including: functional completeness checks to confirm that all necessary functions are fully supported; security threshold compliance to verify whether the final εeff meets the εmax requirement; computational resource feasibility to confirm that the total computational cost and communication overhead are within the system's resource capacity; and protocol consistency verification to ensure that the selected elements can seamlessly collaborate to form a complete protocol.

[0446] A comprehensive verification of the locally optimal element combination is conducted to check whether it satisfies the independence constraints defined in the uncertain matroid model: functional completeness (whether all necessary functions are covered), security compatibility (whether there are security conflicts between elements), and resource feasibility (whether the total resource requirements are within the system capacity). The element combination is confirmed to constitute an effective matroid basis, forming the final protocol minimum weight basis, which serves as the basis for the optimized privacy computing protocol.

[0447] like Figure 3 As shown, the present invention also provides a high-efficiency privacy computing fusion engine system for medical data, comprising:

[0448] The state management module 10 is used to construct a mapping relationship between state identifiers and prefix tree nodes based on the characteristics of multi-source medical data through a prefix tree data structure, and to implement a read-write lock mechanism that supports multi-threaded concurrent access and a dynamic memory allocation strategy, thus obtaining the state management module.

[0449] The adaptive technology selection module 20 is used to analyze the performance characteristics of various privacy computing technologies in different medical data scenarios according to the state management module, classify computing tasks according to data scale, computing complexity and privacy requirements and model them as optimization problems, design approximate algorithms to integrate data types, computing tasks and system resource states, and generate an adaptive privacy computing technology combination strategy.

[0450] The protocol optimization module 30 is used to decompose the privacy computing protocol into basic operation elements and construct a mathematical model according to the adaptive privacy computing technology combination strategy, define the multi-dimensional weights of the computing cost, communication overhead and security strength of the protocol elements, design a solution algorithm to select the minimum cost protocol element combination that satisfies the preset security threshold, and obtain the optimized privacy computing protocol framework.

[0451] The encryption / decryption module 40 is used to perform privacy sensitivity analysis and classification on medical data based on the optimized privacy computing protocol framework, design a hybrid encryption strategy and implement a distributed key generation, distribution and management system to obtain a medical data encryption / decryption system.

[0452] Privacy computing execution engine 50 is used to standardize and encapsulate various privacy computing technologies and define a unified interface based on the medical data encryption and decryption system, and build an execution engine to realize the collaborative scheduling and execution of multiple technologies, thereby obtaining a medical data privacy computing fusion system.

[0453] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for an efficient privacy-preserving computation fusion engine for medical data, characterized in that, include: Based on the characteristics of multi-source medical data, a mapping relationship between state identifiers and prefix tree nodes is constructed using a prefix tree data structure. This implements a read-write lock mechanism and dynamic memory allocation strategy to support multi-threaded concurrent access, resulting in a state management module. The module includes: analyzing the data types, data scale, and access patterns in the multi-source medical data characteristics; analyzing the state space generated during privacy computation to determine state characteristics and distribution patterns; designing a prefix tree structure suitable for privacy computation of medical data, including node representation, path encoding, and tree balancing strategies, resulting in a prefix tree structure design scheme; and based on the prefix tree structure design scheme, constructing a mapping relationship between state identifiers and prefix tree nodes to achieve efficient storage and retrieval of state information, thus obtaining the state... - Prefix tree mapping system; Based on the state-prefix tree mapping system, for scenarios involving federated learning training tasks and homomorphic encryption computation tasks initiated simultaneously by multiple medical institutions, a read-write lock-based concurrency control mechanism is implemented to support simultaneous reading and mutually exclusive writing of prefix tree nodes by multiple threads, and a dynamic memory allocation strategy is implemented to adaptively adjust the memory pool capacity according to the size of the state space, resulting in a concurrency-optimized prefix tree state management module; Based on the concurrency-optimized prefix tree state management module, a path compression technique for the prefix tree is implemented to merge paths with a single child node, and an incremental persistence mechanism is implemented to write only the changed prefix tree nodes to the storage medium, reducing storage overhead and supporting breakpoint recovery of large-scale medical models, resulting in the state management module; Based on the state management module, the performance characteristics of various privacy computing technologies under different medical data scenarios are analyzed. The computing tasks are classified and modeled as optimization problems according to data scale, computing complexity and privacy requirements. An approximate algorithm is designed to integrate data types, computing tasks and system resource states to generate an adaptive privacy computing technology combination strategy. Based on the adaptive privacy computing technology combination strategy, the privacy computing protocol is decomposed into basic operational elements and a mathematical model is constructed. Multidimensional weights for the computational cost, communication overhead, and security strength of the protocol elements are defined. A solution algorithm is designed to select the minimum cost protocol element combination that satisfies a preset security threshold, resulting in an optimized privacy computing protocol framework. This framework includes: based on the federated learning, homomorphic encryption, and secure multi-party computation technologies determined in the adaptive privacy computing technology combination strategy, decomposing the privacy computing protocols corresponding to each technology into basic operational elements for key exchange, data encryption, ciphertext computation, and result decryption; establishing a protocol element library to obtain a protocol element set; and based on the protocol element set, constructing an uncertain matroid model, using the protocol elements as elements in the matroid, and then... The independence constraints of differential privacy protection strength and computational efficiency requirements for protocol execution are used as matroids to obtain a matroid mathematical model. Based on this model, computational cost (including CPU cycles), communication overhead (including data transmission volume), and security strength (including differential privacy parameter ε) are defined for each protocol element. A comprehensive evaluation system is established to obtain an element weight matrix. Based on this weight matrix, a minimum weight basis algorithm is designed and implemented to select protocol element combinations with differential privacy parameter ε values ​​less than a preset security threshold and the lowest total computational cost and communication overhead, thus obtaining the minimum weight basis. Based on this minimum weight basis, a protocol optimization framework is constructed to achieve dynamic combination and formal verification of protocol elements, resulting in the optimized privacy computation protocol framework. Based on the optimized privacy computing protocol framework, privacy sensitivity analysis and classification of medical data are performed, a hybrid encryption strategy is designed, and a distributed key generation, distribution and management system is implemented to obtain a medical data encryption and decryption system. Based on the aforementioned medical data encryption and decryption system, various privacy computing technologies are standardized and encapsulated, and a unified interface is defined. An execution engine is built to achieve collaborative scheduling and execution of multiple technologies, resulting in a medical data privacy computing fusion system.

2. The method according to claim 1, characterized in that, The analysis examines the performance characteristics of various privacy-preserving computing technologies in different medical data scenarios. Computational tasks are categorized and modeled as optimization problems based on data scale, computational complexity, and privacy requirements. An approximate algorithm is designed to integrate data types, computational tasks, and system resource states to generate an adaptive privacy-preserving computing technology combination strategy, including: Based on the state management module, the state data of each privacy computing technology during the historical execution process is obtained, and the execution time, memory usage and computational accuracy characteristics of federated learning, homomorphic encryption and secure multi-party computation under different medical data types and computing scenarios are analyzed to obtain a technology characteristic evaluation matrix. Based on the aforementioned technical characteristic evaluation matrix, medical data computing tasks are classified according to data scale, computational complexity, and privacy requirements, establishing a task classification system and obtaining a medical computing task category model. Based on the aforementioned medical computing task category model, the privacy computing technology selection problem is modeled as a non-uniform k-center problem, and a cost function and resource constraints are defined to obtain a mathematical optimization model. Based on the mathematical optimization model, an approximate algorithm for minimizing resources is designed and implemented to solve the non-uniform k-center problem, resulting in a technology selection algorithm. Based on the aforementioned technology selection algorithm, taking into account data type, computational task, and system resource status, the optimal privacy computing technology combination strategy is dynamically generated, resulting in the adaptive privacy computing technology combination strategy.

3. The method according to claim 1, characterized in that, A medical data encryption and decryption system is obtained by performing privacy sensitivity analysis and classification on medical data, designing a hybrid encryption strategy, and implementing a distributed key generation, distribution, and management system, including: Based on the optimized privacy computing protocol framework, a privacy sensitivity quantification analysis is performed on patient identity information, diagnostic records and genetic data in medical data. The data is divided into three levels of sensitivity from high to low: high sensitivity, medium sensitivity and low sensitivity. A multi-level protection strategy is established to obtain a data sensitivity classification model. Based on the data sensitivity classification model, a hybrid encryption strategy combining symmetric and asymmetric encryption is designed. Different encryption strengths are selected for data with different sensitivities to obtain the hybrid encryption strategy. Based on the hybrid encryption strategy and the computational cost constraints in the optimized privacy computing protocol framework, the key expansion rounds and block mode parameters of the encryption algorithm are optimized to obtain the optimized encryption algorithm set. Based on the optimized set of encryption algorithms, a distributed key generation, distribution and management system is designed and implemented to ensure key security in a multi-organization collaborative environment, thus obtaining a key management system. Based on the key management system, the encryption and decryption functions are encapsulated into a unified interface and service, supporting calls from various privacy computing modules, thus obtaining the medical data encryption and decryption system.

4. The method according to claim 1, characterized in that, The system standardizes and encapsulates various privacy-preserving computation technologies, defines a unified interface, and constructs an execution engine to achieve collaborative scheduling and execution of multiple technologies, resulting in a medical data privacy-preserving computation fusion system, including: Based on the API interface provided by the medical data encryption and decryption system, federated learning, homomorphic encryption, and secure multi-party computation are standardized and encapsulated. The data input and output of each technology are connected to the encryption and decryption service of the medical data encryption and decryption system. A unified interface is defined, including a data encryption interface, a ciphertext computation interface, and a result decryption interface, to obtain a standardized privacy computation module set. Based on the standardized privacy computing module set, a parsing system for medical data analysis and model training tasks is implemented. The system receives user-submitted requests for disease prediction model training, multi-center clinical data statistical analysis, and genomic data association analysis. It calls the medical data encryption and decryption system to encrypt the original medical data and decomposes the requests for disease prediction model training, multi-center clinical data statistical analysis, and genomic data association analysis into a sequence of privacy computing sub-tasks to obtain a task execution plan. Based on the task execution plan, a privacy computing execution engine core is constructed, integrating the key management function of the medical data encryption and decryption system to achieve multi-technology collaborative scheduling and execution, supporting data operations in encrypted state, thus obtaining the execution engine core system; Based on the core system of the execution engine, a computing resource monitoring and performance analysis system is implemented to monitor the encryption and decryption overhead and key management status of the medical data encryption and decryption system, support runtime technology switching and parameter tuning, and obtain a self-optimizing execution environment. Based on the self-optimizing execution environment, an application programming interface and a result visualization system are designed and implemented. The medical data encryption and decryption system is called to decrypt the calculation results and then display them, thus obtaining the medical data privacy calculation fusion system.

5. The method according to claim 1, characterized in that, The process involves analyzing the state space generated during privacy computation to determine state characteristics and distribution patterns. A prefix tree structure suitable for privacy computation of medical data is designed, including node representation, path encoding, and tree balancing strategies. This results in a prefix tree structure design scheme, including: Based on the data type, data scale and access pattern of the multi-source medical data, the state characteristics and distribution patterns in the state space of medical data are analyzed, the number of states, the frequency of state transitions and the state dependencies are determined, and a medical data state space characteristic model is obtained. Based on the state space feature model of the medical data, a prefix tree node structure is designed, including a node data field, a child node pointer array, and a compression tag. A variable-length encoding scheme is designed for the path encoding, and an AVL tree balancing strategy is designed to maintain the query efficiency of the prefix tree, thus obtaining the prefix tree structure design scheme.

6. The method according to claim 2, characterized in that, Based on the medical computing task category model, the privacy computing technology selection problem is modeled as a non-uniform k-center problem, and a cost function and resource constraints are defined to obtain a mathematical optimization model, including: Based on the medical computing task category model, all classified medical data computing tasks are extracted to form a medical data computing task set. Each task in the medical data computing task set is defined as a feature vector containing data scale, computational complexity, and privacy requirements. Privacy computing technologies, including federated learning, homomorphic encryption, and secure multi-party computation, are defined as a technology set, resulting in a task-technology mapping framework. Based on the task-technology mapping framework, a distance function is defined to represent the cost of executing a task in the medical data computing task set using a technology in the technology set. A capacity constraint is defined for each technology in the technology set to represent the maximum number of tasks that each technology can process simultaneously, and a radius constraint is defined to represent the maximum acceptable execution cost of each technology, thus obtaining a set of constraints. Based on the set of constraints, the objective function is constructed to minimize the largest task-technology distance among all tasks. The constraints are set as follows: the number of tasks assigned to each technology does not exceed the corresponding capacity constraint, and the distance between each task and the corresponding assigned technology does not exceed the radius constraint of the corresponding technology, thus obtaining the mathematical optimization model.

7. The method according to claim 1, characterized in that, The algorithm for finding the minimum weight basis is designed and implemented based on the element weight matrix. This algorithm selects a combination of protocol elements from the protocol elements whose differential privacy parameter ε is less than a preset security threshold and whose total computational cost and communication overhead are minimized, thus obtaining the protocol minimum weight basis. This includes: Based on the computational cost, communication overhead, and security strength weight of each protocol element defined in the element weight matrix, a protocol element selection set is created by extracting all protocol elements in the element weight matrix and sorting all protocol elements in descending order of security strength to obtain a candidate element sequence. Based on the candidate element sequence, a greedy strategy is used to sequentially select protocol elements that satisfy the differential privacy parameter ε value being less than a preset security threshold. The total computational cost and communication overhead of the currently selected element combination are calculated to obtain the initial protocol element combination. Based on the initial combination of protocol elements, element replacement and combination optimization are performed iteratively. Under the premise of ensuring that the differential privacy parameter ε meets the security threshold, the total computational cost and communication overhead are continuously reduced. When the total computational cost and communication overhead cannot be reduced in multiple consecutive iterations, the iteration stops, and a local optimal combination of elements is obtained. Based on the locally optimal combination of elements, the independence constraints and matroid properties of the selected protocol elements are verified to ensure that the combination of elements meets the requirements of protocol integrity and security, and the minimum weight basis of the protocol is obtained.

8. A high-efficiency privacy-preserving computing fusion engine system for medical data, characterized in that, include: The state management module, based on the characteristics of multi-source medical data, constructs a mapping relationship between state identifiers and prefix tree nodes using a prefix tree data structure. It implements a read-write lock mechanism and dynamic memory allocation strategy to support multi-threaded concurrent access. The state management module includes: analyzing the data types, data scale, and access patterns in the multi-source medical data characteristics; analyzing the state space generated during privacy computation to determine state characteristics and distribution patterns; designing a prefix tree structure suitable for privacy computation of medical data, including node representation, path encoding, and tree balancing strategies; and based on the prefix tree structure design scheme, constructing a mapping relationship between state identifiers and prefix tree nodes to achieve efficient storage and retrieval of state information. A state-prefix tree mapping system is obtained. Based on the state-prefix tree mapping system, for scenarios involving federated learning training tasks and homomorphic encryption computation tasks initiated simultaneously by multiple medical institutions, a read-write lock-based concurrency control mechanism is implemented to support simultaneous reading and mutually exclusive writing of prefix tree nodes by multiple threads. A dynamic memory allocation strategy is implemented to adaptively adjust the memory pool capacity according to the size of the state space, resulting in a concurrency-optimized prefix tree state management module. Based on the concurrency-optimized prefix tree state management module, a path compression technique for the prefix tree is implemented to merge paths with a single child node. An incremental persistence mechanism is implemented to write only the changed prefix tree nodes to the storage medium, reducing storage overhead and supporting breakpoint recovery of large-scale medical models, resulting in the state management module. The adaptive technology selection module is used to analyze the performance characteristics of various privacy computing technologies in different medical data scenarios based on the state management module, classify computing tasks according to data scale, computing complexity and privacy requirements and model them as optimization problems, design approximate algorithms to integrate data types, computing tasks and system resource states, and generate an adaptive privacy computing technology combination strategy. The protocol optimization module is used to decompose the privacy computing protocol into basic operational elements and construct a mathematical model based on the adaptive privacy computing technology combination strategy. It defines multi-dimensional weights for the computational cost, communication overhead, and security strength of the protocol elements, designs a solution algorithm to select the minimum cost protocol element combination that satisfies a preset security threshold, and obtains the optimized privacy computing protocol framework. This includes: based on the federated learning, homomorphic encryption, and secure multi-party computation technologies determined in the adaptive privacy computing technology combination strategy, decomposing the privacy computing protocols corresponding to each technology into basic operational elements for key exchange, data encryption, ciphertext computation, and result decryption; establishing a protocol element library to obtain a protocol element set; and constructing an uncertain matroid model based on the protocol element set, using the protocol elements as elements in the matroid. The matroid mathematical model is obtained by using the differential privacy protection strength (ensuring data integrity) and the computational efficiency requirements of protocol execution as independence constraints. Based on this model, a comprehensive evaluation system is established for each protocol element, defining computational cost (including CPU cycles), communication overhead (including data transmission volume), and security strength (including the differential privacy parameter ε value), resulting in an element weight matrix. A minimum weight basis algorithm is designed and implemented based on this matrix to select protocol elements whose differential privacy parameter ε value is less than a preset security threshold and whose total computational cost and communication overhead are minimized, thus obtaining the minimum weight basis. Based on this minimum weight basis, a protocol optimization framework is constructed to achieve dynamic combination and formal verification of protocol elements, resulting in the optimized privacy computation protocol framework. The encryption / decryption module is used to perform privacy sensitivity analysis and classification on medical data based on the optimized privacy computing protocol framework, design a hybrid encryption strategy, and implement a distributed key generation, distribution, and management system to obtain a medical data encryption / decryption system. A privacy computing execution engine is used to standardize and encapsulate various privacy computing technologies and define a unified interface based on the medical data encryption and decryption system. The execution engine is then used to achieve collaborative scheduling and execution of multiple technologies, resulting in a medical data privacy computing fusion system.

Images