Heterogeneous computing resource capability radiation scheduling method and device

Abstract

The invention relates to a heterogeneous computing resource capability radiation scheduling method and device. The method comprises the following steps: determining computing capabilities of a plurality of receiving nodes in a radiation center; splitting the original encryption calculation task into a plurality of privacy protection calculation task packages; determining a node matching result according to the computing capabilities of the plurality of receiving nodes and the Pareto optimal solutions of the plurality of privacy protection computing task packages; distributing the plurality of privacy protection computing task packets to corresponding receiving nodes to obtain a distribution voucher of each receiving node; performing distributed verification on the distribution voucher to obtain a distributed verification result; performing multi-modal aggregation on all the distributed verification results to obtain an encrypted calculation radiation scheduling result; the problems of data privacy leakage, low resource utilization rate and insufficient task collaboration efficiency in a traditional scheduling mechanism are solved through encryption task disassembly, dynamic node matching capability, distributed verification and multi-modal aggregation, and the method has the advantages of improving data privacy, resource utilization rate and task collaboration efficiency.

Description

Technical Field

[0001] This invention relates to the field of heterogeneous resource scheduling technology, and in particular to a method and apparatus for heterogeneous computing resource capacity radiation scheduling. Background Technology

[0002] With the continuous development of computer technology, cloud computing, as one of the most promising technologies, has gradually become the preferred solution for most enterprises to solve IT problems, and it has significant advantages in optimizing business processes, reducing costs, and improving system security. However, in the process of implementing cloud computing, the management of heterogeneous resources has become a challenge. Modern applications (such as AI training and scientific computing) often involve mixed tasks such as parallel computing, serial logic, and real-time processing, which require the collaborative completion of heterogeneous resources.

[0003] Traditional heterogeneous computing resource scheduling suffers from problems such as data privacy leaks, uneven resource allocation, and low computational efficiency. While existing technologies such as secure multi-party computation and federated learning can protect data privacy to some extent, they cannot fully utilize the advantages of heterogeneous computing resources, resulting in the inability to achieve unified scheduling of mixed tasks.

[0004] Therefore, there is an urgent need to propose a heterogeneous computing resource capacity radiation scheduling method and device to solve the technical problem that existing heterogeneous computing resource scheduling methods cannot fully utilize the advantages of heterogeneous computing resources, resulting in the inability to achieve unified scheduling of mixed tasks. Summary of the Invention

[0005] In view of this, it is necessary to provide a heterogeneous computing resource capacity radiation scheduling method and apparatus to solve the technical problem that existing heterogeneous computing resource scheduling methods cannot fully utilize the advantages of heterogeneous computing resources, resulting in the inability to achieve unified scheduling of mixed tasks.

[0006] To address the aforementioned problems, in a first aspect, the present invention provides a heterogeneous computing resource capacity radiation scheduling method, comprising: Obtain the original encrypted computation task and confirm the computing capabilities of multiple receiving nodes in the radiation center; The original encrypted computing task is broken down into multiple privacy-preserving computing task packages; Based on the computing power of the plurality of receiving nodes and the Pareto optimal solution of the plurality of privacy-preserving computing task packages, the node matching result is confirmed; Based on the node matching results, the multiple privacy-preserving computing task packages are allocated to the corresponding receiving nodes, and the allocation credential for each receiving node is obtained; The allocation credential is subjected to distributed verification to obtain the distributed verification result; Multimodal aggregation of all distributed verification results yields encrypted computation radiation scheduling results.

[0007] In one possible implementation, confirming the computing power of multiple receiving nodes in the radiation center includes: Acquire basic data and historical reliability records of the multiple receiving nodes at the radiation center; the basic data includes data in the dimensions of computation, storage, security, and network. A multidimensional capability profile for each receiving node is constructed based on the data from the computational dimension, the storage dimension, the security dimension, and the network dimension. The corresponding computing capabilities are obtained based on the multidimensional capability profile and historical reliability record of each receiving node.

[0008] In one possible implementation, the metadata portion of the original cryptographic computation task includes structured and unstructured data; the step of decomposing the original cryptographic computation task into multiple privacy-preserving computation task packages includes: After decrypting the metadata part of the original encrypted computing task, the fields in the structured data are encrypted according to the preset sensitivity of the fields to obtain structured encrypted data. Then, the structured encrypted data is subjected to multi-dimensional feature clustering to obtain multiple feature information. The unstructured data is detected and segmented according to a preset target detection model to obtain segmented region information; The multiple feature information and the block region information are stored in multiple segments of the radiation center to obtain a mapping table for each segment; The original encryption computation task is converted into a directed acyclic graph, and the directed acyclic graph is decomposed to obtain the decomposition result; Based on the decomposition results and the mapping table, multiple initial privacy-preserving computation task packages are generated; After generating a unique privacy credential based on the data in the fragments of each initial privacy-preserving computation task package, four checks are performed on each initial privacy-preserving computation task package based on the unique privacy credential to obtain the check results; the four checks include data integrity, logical completeness, privacy compliance, and performance prediction. The initial privacy-preserving computation task package that passes the inspection is marked as ready, resulting in multiple privacy-preserving computation task packages marked as ready.

[0009] In one possible implementation, after decomposing the original cryptographic computation task into multiple privacy-preserving computation task packages, the method further includes: After the original encrypted computing task is broken down by bandwidth, if the bandwidth decreases by a preset percentage, the fragments in the multiple privacy-preserving computing task packages are split and recombined to obtain multiple privacy-preserving computing task packages with smaller capacity.

[0010] In one possible implementation, confirming the node matching result based on the computing power of the plurality of receiving nodes and the Pareto optimal solution of the plurality of privacy-preserving computing task packages includes: The multiple receiving nodes are evaluated based on the evaluation indicators and the computing power to obtain the evaluation results; The execution order constraints of the multiple privacy-preserving computation task packages are determined based on the dependencies in the directed acyclic graph. Based on the NSGA-II algorithm, the Pareto optimal solution for each privacy-preserving computation task package is determined among the four checks of the multiple privacy-preserving computation task packages in terms of privacy protection strength, computational efficiency, and cost. The node matching result is confirmed based on the execution order constraint, the Pareto optimal solution, and the evaluation result.

[0011] In one possible implementation, the step of allocating the plurality of privacy-preserving computation task packages to corresponding receiving nodes based on the node matching results, and obtaining the allocation credential for each receiving node, includes: Set up encrypted transmission channels and fault tolerance mechanisms; Each privacy-preserving computing task package is transmitted to the corresponding receiving node through the dynamic topology network, the encrypted transmission channel, and the node matching result. The transmission process is monitored according to the fault tolerance mechanism to obtain the monitoring result. When the monitoring result is abnormal, the corresponding recovery process is initiated according to the fault type in the monitoring result, so that each privacy protection computing task package is transmitted to the corresponding receiving node; When the receiving node finishes receiving, it obtains the allocation credential for each receiving node.

[0012] In one possible implementation, the distributed verification of the allocation credential to obtain a distributed verification result includes: Based on the allocation credentials, cross-validation is performed on the node groups that have completed the same privacy-preserving computation task package to obtain the cross-validation results; Randomly select at least one privacy-preserving computation task package for re-execution and spot check verification to obtain the integrity verification result; A privacy compliance audit was performed on the multiple privacy-preserving computing task packages, and the audit results were obtained. The logical consistency of the multiple privacy-preserving computation task packages is verified based on the directed acyclic graph to obtain the global dependency verification results. A distributed verification result is generated based on threshold cryptography, the cross-validation result, the integrity verification result, the audit result, and the global dependency verification result.

[0013] In one possible implementation, the step of multimodal aggregation of all distributed verification results to obtain the cryptographic computation radiation scheduling result includes: Integrity verification is performed on all privacy-preserving computation task packages that participate in the aggregation of all distributed verification results to obtain the verification result; After cleaning the data of the multiple privacy-preserving computing task packages based on the verification results, the multiple privacy-preserving computing task packages are then aggregated in three levels to obtain the final aggregation result; the three-level aggregation includes node level, computing cluster level and aggregation level. The final aggregation result is evaluated in four dimensions to obtain the encrypted computation radiation scheduling result; the four-dimensional evaluation includes logical consistency, statistical rationality, computational correctness and privacy compliance.

[0014] In one possible implementation, the three-level aggregation of the multiple privacy-preserving computation task packages to obtain the final aggregation result includes: The multiple privacy-preserving computation task packages are aggregated at three levels to obtain multi-level result alignment results; Differential privacy protection is dynamically injected into the alignment results of the multi-level results to obtain privacy-enhancing processing results; The privacy-enhancing processing results are encrypted and aggregated using a secure multi-party computation protocol to obtain the multi-party computation aggregation result; The multi-party computation aggregation results are subjected to anomaly detection and conflict resolution to obtain the final aggregation result.

[0015] Secondly, the present invention also provides a heterogeneous computing resource capacity radiation scheduling device, comprising: The task receiving module is used to acquire the original encrypted computing task and confirm the computing capabilities of multiple receiving nodes in the radiation center. The task package decomposition module is used to decompose the original encrypted computing task into multiple privacy-preserving computing task packages; The node matching module is used to confirm the node matching result based on the computing power of the plurality of receiving nodes and the Pareto optimal solution of the plurality of privacy-preserving computing task packages; The task package allocation module is used to allocate the multiple privacy-preserving computing task packages to the corresponding receiving nodes according to the node matching results, and obtain the allocation credential for each receiving node; The allocation verification module is used to perform distributed verification on the allocation credential and obtain the distributed verification result. The multimodal aggregation module is used to aggregate all distributed verification results in a multimodal manner to obtain the encrypted computing radiation scheduling result.

[0016] The beneficial effects of this invention are as follows: It obtains the original encrypted computing task and confirms the computing capabilities of multiple receiving nodes in the radiation center; it decomposes the original encrypted computing task into multiple privacy-preserving computing task packages; based on the computing capabilities of multiple receiving nodes and the Pareto optimal solutions of multiple privacy-preserving computing task packages, it confirms the node matching results; based on the node matching results, it allocates multiple privacy-preserving computing task packages to the corresponding receiving nodes, obtaining an allocation credential for each receiving node; it performs distributed verification on the allocation credentials to obtain distributed verification results; it performs multimodal aggregation on all distributed verification results to obtain the encrypted computing radiation scheduling results; by decomposing the encrypted task, dynamically matching node capabilities, performing distributed verification, and multimodal aggregation, it solves the problems of data privacy leakage, low resource utilization, and insufficient task collaboration efficiency in traditional scheduling mechanisms, and has the advantages of improving data privacy, resource utilization, and task collaboration efficiency. Attached Figure Description

[0017] Figure 1 A schematic flowchart of an embodiment of the heterogeneous computing resource capacity radiation scheduling method provided by the present invention; Figure 2 For the present invention Figure 1 A schematic diagram of an embodiment of step S101; Figure 3 For the present invention Figure 1 A schematic diagram of an embodiment of step S102; Figure 4 For the present invention Figure 1 A schematic flowchart of an embodiment of step S103; Figure 5 For the present invention Figure 1 A schematic flowchart of an embodiment of step S105; Figure 6 For the present invention Figure 1 A schematic flowchart of an embodiment of step S106; Figure 7 For the present invention Figure 6 A schematic flowchart of an embodiment of step S602; Figure 8 A schematic diagram of an embodiment of the target adversarial patch defense model provided by the present invention for testing; Figure 9 This is a schematic diagram of an embodiment of the heterogeneous computing resource capacity radiation scheduling device provided by the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] Privacy-preserving computation refers to an information technology that analyzes and computes data without disclosing the original data, ensuring that the data provider does not reveal the original data. It guarantees that data is "usable but not visible" throughout its entire flow, from generation and storage to computation, application, and destruction. Common technical solutions for privacy-preserving computation include secure multi-party computation, federated learning, trusted execution environments, and encrypted computation. Common underlying technologies include obfuscated circuits, unintended transmission, secret sharing, and homomorphic encryption.

[0020] Capability radiation scheduling: By leveraging the properties of quantum entanglement and superposition, quantum computing resources (QPU) and traditional heterogeneous computing resources (CPU / GPU / FPGA) are dynamically allocated to achieve ultra-parallel execution of computing tasks and enhance privacy.

[0021] Entropy Hidden Architecture: A computational framework designed based on the principle of quantum entropy reduction, which achieves synergistic optimization of data privacy and computational efficiency through quantum noise suppression and classical-quantum hybrid coding techniques.

[0022] QCU: Quantum Computing Unit, a hardware module that supports the collaboration between quantum state operations and classical computing.

[0023] PPCU: Privacy-Preserving Computing Unit.

[0024] like Figure 1 As shown, a specific embodiment of the present invention discloses a heterogeneous computing resource capacity radiation scheduling method, comprising: S101. Obtain the original encryption computation task and confirm the computing capabilities of multiple receiving nodes in the radiation center.

[0025] The original encrypted computing task refers to the encryption processing requirements involving mixed data types. Specifically, it can be implemented using layered encryption technology, performing field-level encryption on structured data and block-level encryption on unstructured data. In the entropy-hidden architecture, the radiation center is the core scheduling unit that dynamically allocates PPCUs (Privacy-Preserving Computation Packages) based on the computing capabilities of the receiving nodes. The radiation center is a key node in the network topology, responsible for collecting real-time capabilities (such as computing power, memory, bandwidth, etc.) of the receiving nodes and serving as the decision point for resource allocation. The radiation center plays the role of an "intelligent allocator" in the entropy-hidden architecture; its core value lies in achieving efficient and stable system operation under uncertain environments through dynamic, non-uniform resource radiation allocation. The receiving nodes are the actual bearers of the PPCUs, receiving resource quotas dynamically allocated by the radiation center and executing specific computing tasks. Each receiving node needs to report its own capability parameters (such as CPU / GPU computing power, memory bandwidth, network I / O, etc.) in real time; this data constitutes the basis for the radiation center's decision to allocate PPCUs.

[0026] In this embodiment of the invention, it can be applied to an entropy-hidden architecture. Within this architecture, a series of subsequent processes can be performed. The system receives a task container (i.e., the original encrypted computation task) encrypted with AES-256. This container contains three parts: original data, a computational logic description, and privacy parameters. In the SGX trusted execution environment, only the metadata portion of the original data (including task type, computational flowchart, and privacy budget) is decrypted, while the core data remains encrypted, ensuring that sensitive information is not exposed throughout the decomposition process. The decryption process can employ two-factor authentication, such as simultaneously verifying the platform's hardware fingerprint and the task issuer's digital signature.

[0027] S102. Decompose the original encrypted computing task into multiple privacy-preserving computing task packages.

[0028] Privacy-preserving computation task packages refer to task units divided according to data sensitivity and processing requirements. Specifically, this can be achieved through directed acyclic graph decomposition combined with fragmented storage technology, ensuring logical independence between task packages. Based on predefined data classification standards (such as HIPAA medical data grading), a sensitivity score can be assigned to each data field, thereby calculating the sensitivity and decomposing the original encrypted computation task into multiple privacy-preserving computation task packages.

[0029] S103. Based on the computing power of multiple receiving nodes and the Pareto optimal solution of multiple privacy-preserving computing task packages, confirm the node matching result.

[0030] Pareto optimality refers to finding a set of non-dominated solutions in multi-objective optimization, which can be achieved using the NSGA-II algorithm, balancing privacy strength, computational efficiency, and resource cost. Thus, the optimal solution for each privacy-preserving computational task package can be determined through Pareto optimality, and the optimal solution and receiving nodes can be matched based on the computational capabilities of multiple receiving nodes to confirm the node matching results.

[0031] S104. Based on the node matching results, allocate multiple privacy-preserving computing task packages to the corresponding receiving nodes and obtain the allocation credential for each receiving node.

[0032] The allocation credential refers to the encrypted credentials for task allocation, which can be implemented using blockchain notarization technology and includes task hash and timestamp information. The node matching result can include the receiving nodes to be assigned to each privacy-preserving computation task package. Therefore, multiple privacy-preserving computation task packages can be allocated to the corresponding receiving nodes based on the node matching result. When the allocation is complete, the allocation credential for each receiving node can be obtained.

[0033] S105. Perform distributed verification on the allocation credential to obtain the distributed verification result.

[0034] Distributed verification refers to a verification mechanism involving multiple parties, which can be implemented using threshold signature technology and cross-validation to ensure data integrity. Distributed verification enables multi-faceted checks, guaranteeing data accuracy.

[0035] S106. Perform multimodal aggregation on all distributed verification results to obtain the encrypted computing radiation scheduling results.

[0036] Multimodal aggregation refers to the fusion processing of multidimensional data, which can be achieved by using a secure multi-party computation protocol to eliminate single-point verification bias.

[0037] Specifically, the system first scans available computing nodes (i.e., receiving nodes) to construct a three-layer radial network topology (i.e., a dynamic topology network). The core layer (L1) consists of high-performance servers with TEE security environment, connected via fiber optic leased lines of 10Gbps or higher; the edge layer (L2) consists of edge computing nodes (i.e., receiving nodes) distributed in various locations, with bandwidth maintained at 1-10Gbps; the terminal layer (L3) connects to various network devices, and the network status changes dynamically. The topology maintenance module checks the node liveness status every 30 seconds using ICMP+TCP dual probes and updates the link quality matrix between computing nodes (i.e., receiving nodes) in real time. This matrix includes 12 network metrics such as bandwidth, latency, and packet loss rate. The specific network topology construction process can be configured according to actual conditions, and this embodiment of the invention does not impose any limitations. The dynamic topology network can provide bandwidth within the entropy-hidden architecture, and all processes are completed based on the dynamic topology network. When an encrypted computing task sent by a user terminal (which can be configured according to actual conditions, such as a computer) is received through the dynamic topology network, the task metadata is first parsed to identify the structured and unstructured parts. Structured data fields are encrypted with sensitivity-level classification, forming sub-task packages after feature clustering; unstructured data is segmented for target detection, generating independent processing units. The granularity of task packages is dynamically adjusted based on real-time network bandwidth status, automatically splitting packages to ensure transmission efficiency when bandwidth decreases. A multi-dimensional node capability evaluation model is established, combining historical reliability data to select qualified nodes. A multi-objective optimization algorithm is used to match the optimal node combination for each task package, generating encrypted transmission channels and implementing transmission monitoring. During task execution, random re-execution checks and logical consistency verification ensure reliable processing. In the final aggregation stage, differential privacy technology is used to process intermediate results, and an anomaly detection mechanism eliminates the impact of erroneous data.

[0038] Compared with existing technologies, this embodiment provides a method for obtaining the original encrypted computing task and confirming the computing capabilities of multiple receiving nodes in the radiation center; decomposing the original encrypted computing task into multiple privacy-preserving computing task packages; confirming the node matching results based on the computing capabilities of multiple receiving nodes and the Pareto optimal solutions of multiple privacy-preserving computing task packages; allocating multiple privacy-preserving computing task packages to corresponding receiving nodes according to the node matching results, obtaining an allocation credential for each receiving node; performing distributed verification on the allocation credential to obtain distributed verification results; and performing multimodal aggregation on all distributed verification results to obtain the encrypted computing radiation scheduling result. By decomposing the encrypted task, dynamically matching node capabilities, performing distributed verification, and multimodal aggregation, this method solves the problems of data privacy leakage, low resource utilization, and insufficient task collaboration efficiency in traditional scheduling mechanisms, and has the advantages of improving data privacy, resource utilization, and task collaboration efficiency.

[0039] Through the above technical solutions, this application achieves precise matching between encrypted computing tasks and heterogeneous resources, improving resource utilization while ensuring data privacy. The dynamic task package adjustment mechanism effectively copes with network fluctuations, avoiding overall task stagnation due to transmission delays. The distributed verification system ensures that the processing at each stage is traceable and verifiable, preventing single points of failure from affecting task reliability. The multi-level aggregation mechanism eliminates the impact of local errors on the final result, providing secure and efficient scheduling support for hybrid computing tasks.

[0040] In some embodiments of the present invention, such as Figure 2 As shown, step S101 includes: Step S201: Obtain basic data and historical reliability records of multiple receiving nodes at the radiation center; the basic data includes data in the computational dimension, storage dimension, security dimension, and network dimension.

[0041] The computing dimension refers to the computational capability of the receiving node when processing tasks. This can be quantified using floating-point operation speed or the number of parallel threads, and is used to measure the node's ability to execute computationally intensive tasks. The storage dimension refers to the capacity and access speed of the receiving node's local storage resources, assessed for example by testing random read / write speed or maximum storage space, and is used to determine the node's ability to handle data storage requirements. The security dimension refers to the encryption mechanisms used by the receiving node during data transmission and storage, assessed for example by checking whether it supports homomorphic encryption or a trusted execution environment, and is used to ensure privacy protection during task execution. The network dimension refers to the communication quality between the receiving node and the radiation center, specifically measured using bandwidth testing tools to measure transmission rate and latency, and is used to optimize network efficiency during task allocation. Historical reliability records refer to the reliability data demonstrated by the receiving node in past task executions, generated for example by statistically analyzing task completion rates or error frequencies, and are used to dynamically adjust node priority weights.

[0042] Step S202: Construct a multi-dimensional capability profile for each receiving node based on data from the computational, storage, security, and network dimensions.

[0043] The multidimensional capability profile refers to a structured evaluation model formed by integrating data from four dimensions. For example, the entropy weight method is used to weight different dimensional indicators to comprehensively reflect the actual processing capabilities of a node. A multidimensional capability profile is created for each receiving node and each registered node, including: computing dimension (CPU architecture / GPU model / FPGA logic unit count), storage dimension (memory capacity / SSD IOPS), security dimension (TEE type / certification certificate), and network dimension (uplink / downlink bandwidth / NAT type). Actual computing power indicators are measured through benchmark testing programs. This data is normalized and stored in a graph database to form a node knowledge graph containing 200+ feature dimensions.

[0044] Step S203: Obtain the corresponding computing capabilities based on the multidimensional capability profile and historical reliability records of each receiving node.

[0045] Specifically, when confirming the computing capabilities of receiving nodes, the basic operating parameters of each node are first obtained from the radiation center, including computing capability test results, storage performance reports, security configuration information, and network connection status. For example, computing dimension data can be obtained through benchmark testing tools, storage dimension data through disk performance analysis tools, security dimension data by scanning the node's encryption protocol support, and network dimension data by real-time monitoring of transmission link quality. Subsequently, the raw data from these four dimensions are normalized and combined with credibility indicators such as node response time and task interruption counts recorded in historical task execution records to construct a multi-dimensional capability profile for each node. This multi-dimensional capability profile is then integrated with historical credibility records; for example, evaluation metrics are used to assess the data of each receiving node, outputting a quantitative score reflecting the node's overall computing capability.

[0046] Through the above technical solution, this application can accurately evaluate the comprehensive processing capabilities of heterogeneous nodes, taking into account multiple factors such as computing efficiency, storage capacity, security protection and network transmission when allocating tasks.

[0047] In some embodiments of the present invention, such as Figure 3 As shown, the metadata portion of the original cryptographic computation task includes structured and unstructured data; step S102 includes: S301. After decrypting the metadata part of the original encrypted computing task, the fields in the structured data are encrypted according to the preset sensitivity of the fields to obtain structured encrypted data. Then, multi-dimensional feature clustering is performed on the structured encrypted data to obtain multiple feature information.

[0048] Structured data refers to data with a predefined format, which can be implemented using database tables or JSON format. Encrypting fields with preset sensitivity levels ensures differentiated protection for data at different sensitivity levels. Multidimensional feature clustering refers to grouping data based on multidimensional similarity according to data attributes. This can be implemented using K-means or hierarchical clustering algorithms to optimize the correlation of task packages.

[0049] In this embodiment of the invention, based on predefined data classification standards (such as HIPAA medical data grading), a sensitivity score (level 1-5) is assigned to each data field. Combining the historical trust records of the receiving nodes (stored on the blockchain), the actual required privacy protection strength is dynamically calculated using the formula: Privacy Level = [(Sensitivity Score × Data Volume) / Trust Threshold]. For example, a task containing 100,000 medical records with a sensitivity level of 4, when assigned to a node with a trust level of 80%, will generate a strict protection requirement of privacy level 5. Differentiated processing is implemented according to field sensitivity: Level 5 fields such as ID card numbers are processed using SHA-3 hashing with a random salt value; Level 3-4 fields (such as address, consumption records) are processed using Paillier homomorphic encryption; Level 1-2 fields (such as age range, gender) are discretized into intervals to obtain structured encrypted data after encryption. The processing method for each field can also be recorded in the metadata, for example, converting "Age: 35" to "Age Range: 30-40", while retaining the mapping relationship of the original value in the encrypted state. An improved k-anonymization algorithm is used to perform multidimensional feature clustering on the data (such as diagnosis type + age group + gender) to obtain multiple feature information.

[0050] S302. Detect and segment unstructured data according to the preset target detection model to obtain segmented region information.

[0051] Unstructured data refers to data without a fixed format, specifically in the form of images, text, or video. Key region information can be extracted by segmenting the data using object detection models (such as the YOLO model). First, the YOLO model is used to detect sensitive regions (such as faces and license plates). Gaussian blur (σ=3.0) and local differential privacy (ε=0.5) are applied to these regions, while background regions are segmented as is. Text data is extracted into conceptual entities using the BERT model and then converted into standardized medical codes (e.g., "hypertension" → ICD-10:I10), while a homomorphically encrypted copy of the original text is retained for specific queries.

[0052] S303. Store multiple feature information and block region information in multiple segments of the radiation center to obtain a mapping table for each segment.

[0053] In this context, a shard refers to an independent unit of data storage, which can be implemented using logical partitions in a distributed storage system. Recording data locations through a mapping table improves retrieval efficiency. Multiple feature information of structured data and block information of unstructured data are stored in multiple shards at the radiation center. This ensures that each shard contains at least k indistinguishable records (default k=50; for example, if gender is male, then it cannot be distinguished by female). Differential privacy-based cluster center perturbation technology prevents the inference of individual identity based on shard features. The sharding process generates a data-shard mapping table, which is stored using threshold encryption technology and requires multiple administrators to decrypt.

[0054] S304. Convert the original encryption computation task into a directed acyclic graph, and decompose the directed acyclic graph to obtain the decomposition result.

[0055] In this context, a directed acyclic graph (DAG) refers to a task execution path without cyclic dependencies. Specifically, it can be decomposed using a topological sorting algorithm to clarify the task execution order. The task description is converted into a DAG, and the Tarjan algorithm is used to identify strongly connected components. The input / output dependencies of each computation node are analyzed, and critical path nodes (nodes affecting the final result) are marked. For example, in federated learning tasks, gradient aggregation nodes are identified as critical paths, while local model training nodes can be processed in parallel. The dependency analysis results are stored in JSON format, containing node IDs, execution order, and parallelism information, thus yielding the decomposition results.

[0056] S305. Based on the decomposition results and mapping table, generate multiple initial privacy-preserving computation task packages.

[0057] Based on the DAG decomposition results and mapping table, an independent PPCU package (i.e., the initial privacy-preserving computation task package) is generated for each parallelizable unit. Each PPCU contains: a processed data fragment (accounting for 1 / N of the original data), a subset of computational logic (such as statistical analysis code for specific features, or encrypted computation and processing procedures), and verification metadata (containing Merkle tree position hashes and parent node signatures). The PPCU employs a nested encryption structure: the outer layer is encrypted using the target node's public key, while the inner layer retains a homomorphic encryption state for computation.

[0058] S306. After generating a unique privacy credential based on the data in the fragments of each initial privacy-preserving computation task package, perform four checks on each initial privacy-preserving computation task package based on the unique privacy credential, and obtain the check results; the four checks include data integrity, logical completeness, privacy compliance, and performance prediction.

[0059] Among the four checks, logical completeness refers to the absence of internal logical conflicts within the task package. This can be achieved through static code analysis tools to avoid logical errors during execution. A unique privacy credential is generated for each PPCU, containing: a differential privacy budget allocation record (e.g., this PPCU allows ε=0.1), a data lineage tag (using zk-SNARKs to prove the legitimacy of the data source), and an audit seed generated by a verifiable random function (VRF). This metadata is organized into a Merkle Patricia tree structure, with the root hash value registered on the blockchain for subsequent verification. Finally, four checks are performed on all PPCUs based on the unique privacy credential: data integrity (hash verification), logical completeness (dependency verification), privacy compliance (ε value pre-calculation), and performance estimation (execution time analysis). Monitoring results are then obtained.

[0060] S307. Mark the initial privacy-preserving computation task package that has passed the inspection as ready, and obtain multiple privacy-preserving computation task packages marked as ready.

[0061] Among them, PPCUs that pass the inspection are marked as ready, resulting in multiple privacy-preserving computing task packages marked as ready, which await scheduling according to the computing power radiation allocation strategy. At the same time, a task decomposition report is generated, recording the metadata index and expected execution path of each PPCU.

[0062] Specifically, the metadata portion of the original encrypted computation task is split into structured and unstructured data. The structured data undergoes decryption and secondary encryption of sensitive fields to ensure that highly sensitive information remains protected in subsequent processing. Multidimensional feature clustering divides the encrypted structured data into multiple feature information groups, allowing task packages to be allocated according to the computational characteristics of nodes. Unstructured data is region-divided using object detection models, such as identifying and segmenting key regions in image data to reduce redundant computation. Feature information and segmented region information are distributed and stored across multiple fragments, with an index relationship established through a mapping table for quick data location during subsequent task package generation. The original task is converted into a directed acyclic graph and decomposed into multiple sub-task nodes, ensuring that the task execution order conforms to dependencies. The generated initial task package, after being appended with a unique privacy credential, undergoes sequential data integrity verification, logical completeness verification, privacy compliance review, and performance estimation. For example, hash verification ensures data has not been tampered with, or execution is simulated in a sandbox environment to estimate resource consumption. Task packages that pass these four checks are marked as ready, providing reliable input for subsequent scheduling.

[0063] Through the above technical solutions, this application effectively addresses the issues of privacy protection and task splitting efficiency in processing mixed data types. Sensitivity-level encryption of structured data ensures differentiated protection for different fields, avoiding performance degradation or insufficient protection caused by traditional unified encryption. Dynamic block partitioning of unstructured data, combined with an object detection model, improves the computational priority and resource utilization of key areas. Through directed acyclic graph decomposition and a four-check mechanism, the task package generation process balances execution order constraints and multi-dimensional quality assurance, providing highly reliable input for subsequent node matching and scheduling, thereby reducing task execution failure rates and minimizing the risk of privacy leakage.

[0064] In some embodiments of the present invention, after decomposing the original encryption computation task into multiple privacy-preserving computation task packages, the method further includes: After the original encrypted computing task is broken down by bandwidth, if the bandwidth decreases by a preset percentage, the fragments in the multiple privacy-preserving computing task packages are split and recombined to obtain multiple privacy-preserving computing task packages with smaller capacity.

[0065] In a specific embodiment of the present invention, because the embodiments of the present invention are based on a dynamic topology network, the network status of the bandwidth of the dynamic topology network can be monitored in real time. If, after the original encryption computation task is decomposed into multiple privacy-preserving computation task packages, a bandwidth drop of a preset percentage (e.g., more than 30%) is detected, fragment reassembly is automatically triggered: specifically, the original fragment is split into smaller units (as small as 1kB), and the redundancy strategy is adjusted (from 3 replicas to 5 replicas). At the same time, the computational complexity of fragmentation is dynamically adjusted according to the node CPU load. For example, matrix operations that originally required a single node to complete are split into multiple vector dot product tasks, resulting in multiple privacy-preserving computation task packages with smaller capacity.

[0066] Through the above technical solution, this application can maintain task transmission efficiency in scenarios with unstable network bandwidth, and prevent data packet loss or task interruption caused by sudden bandwidth drops. The split sub-fragments can make full use of unoccupied transmission resources in dynamic topology networks, such as reducing the risk of single point of failure through multi-path parallel transmission, while maintaining the privacy protection strength of the original task packet unaffected by changes in fragment size.

[0067] In some embodiments of the present invention, such as Figure 4 As shown, step S103 includes: S401. Evaluate multiple receiving nodes based on evaluation indicators and computing power to obtain evaluation results.

[0068] The evaluation indicators refer to a multi-dimensional evaluation system used to quantify the comprehensive capabilities of nodes. Specifically, this can be implemented using a weighted scoring model across computational, storage, security, and network dimensions to objectively reflect the actual processing capabilities of nodes. Nodes are screened through a blockchain-based reputation system, with evaluation indicators including: historical task completion rate (40% weight), result accuracy (30%), response latency stability (20%), and security event records (10%). Candidate nodes must provide remote proof reports, and the verification service platform must verify that the configuration meets TEE requirements.

[0069] S402. Determine the execution order constraints of multiple privacy-preserving computation task packages based on the dependencies in the directed acyclic graph.

[0070] The execution order constraint refers to the logical dependencies between task packages, which can be extracted using a topological sorting algorithm for directed acyclic graphs (DAGs) to ensure that the order of task execution conforms to business logic. When a PPCU task package is received, the scheduling engine performs multi-objective optimization decisions: first, it determines the execution order constraints based on the DAG dependencies.

[0071] S403. Based on the NSGA-II algorithm, determine the Pareto optimal solution for each privacy-preserving computation task package among the four checks of privacy protection strength, computational efficiency, and cost.

[0072] The NSGA-II algorithm refers to the second-generation non-dominated sorting genetic algorithm, which can be implemented using a multi-objective optimization framework. It can find the optimal balance among three conflicting objectives: privacy protection strength, computational efficiency, and cost. The improved NSGA-II algorithm seeks the Pareto optimal solution among the three objectives: privacy protection strength (requiring ε≤0.5), computational efficiency (delay SLA), and cost (cost unit).

[0073] S404. Confirm the node matching result based on the execution order constraint, Pareto optimal solution, and evaluation results.

[0074] The receiving nodes can be matched based on the execution order constraint Pareto optimal solution and evaluation results to obtain the node matching results. For GPU-accelerated PPCUs, CUDA compatibility and video memory capacity are additionally checked; for PPCUs containing sensitive data, it is mandatory to allocate L1 nodes with SGX2.0.

[0075] Specifically, during task scheduling, receiving nodes are first comprehensively scored using multi-dimensional capability profiles to form a node capability evaluation matrix. Simultaneously, based on the directed acyclic graph generated from task decomposition, strong dependencies between task packages are extracted to form an execution order constraint set. Then, the NSGA-II algorithm is used to construct a multi-objective optimization model, using the privacy protection strength parameter, estimated computation time, and resource cost of each task package as optimization objectives. Through iterative processes such as population initialization, non-dominated sorting, and congestion calculation, a Pareto front solution set is generated. Finally, combining the node evaluation results and execution order constraints, the node-task matching scheme that satisfies all constraints and has the highest comprehensive score is selected from the Pareto optimal solution set.

[0076] Through the above technical solution, this application effectively solves the technical problem of multi-objective optimization in heterogeneous computing scenarios. While ensuring data privacy and security, it can dynamically adjust the task allocation strategy according to the real-time resource status, which not only avoids resource idleness and waste, but also prevents node overload, and significantly improves the overall scheduling efficiency of complex computing tasks in heterogeneous environments.

[0077] In some embodiments of the present invention, multiple privacy-preserving computation task packages are allocated to corresponding receiving nodes based on node matching results, and an allocation credential for each receiving node is obtained, including: Configure encrypted transmission channels and fault tolerance mechanisms.

[0078] Before task distribution, an end-to-end encrypted channel is established: ECDH-P256 is used for key negotiation, and the session key is rotated every 5 minutes; the data stream is encrypted using AES-256-GCM, and each data packet is appended with an HMAC-SHA256 signature. For large file transfers, an innovative fragmented hybrid encryption scheme is adopted—asymmetric encryption is used for the metadata portion, and symmetric encryption is used for the data blocks, balancing security and transmission efficiency. The fault tolerance mechanism includes a three-level detection mechanism: process level (heartbeat interval <2s), container level (resource utilization threshold), and host level (hardware health status).

[0079] Each privacy-preserving computing task package is transmitted to the corresponding receiving node through a dynamic topology network, encrypted transmission channels, and node matching results. The transmission process is monitored according to a fault-tolerant mechanism to obtain monitoring results.

[0080] In a specific embodiment, the bandwidth of the connected dynamic topology network can be used to transmit the successfully matched privacy-preserving computation task packets from the node matching results to the receiving nodes via an encrypted transmission channel. Each privacy-preserving computation task packet is transmitted to its corresponding receiving node. During the transmission process, a fault-tolerant mechanism can be used to monitor the transmission process and obtain the monitoring results.

[0081] When the monitoring result is abnormal, the corresponding recovery process is initiated according to the fault type in the monitoring result, so that each privacy protection computing task package is transmitted to the corresponding receiving node.

[0082] When an anomaly is detected, a corresponding recovery process can be initiated based on the type of fault detected in the monitoring results: if a node crashes, it will be migrated to a backup node at the same level; if there is a network partition, it will switch to asynchronous processing mode; if there is a data anomaly, it will trigger trusted re-execution verification.

[0083] When the receiving node finishes receiving, it obtains the allocation certificate for each receiving node.

[0084] In a specific embodiment, after task allocation is completed, verifiable allocation credentials are generated, including information such as node identity verification, task hash, and execution time window. These credentials are aggregated, signed, and stored on the blockchain. The audit module periodically checks the consistency between the actual execution and the allocation plan. If a deviation exceeds 5%, an alarm is automatically triggered, and an audit report containing the evidence chain is generated for regulatory review.

[0085] Furthermore, the reinforcement learning-based load balancer continuously monitors the resource utilization of each node, automatically triggering load migration when CPU usage exceeds 80% or memory usage exceeds 90%. An improved consistent hashing algorithm ensures that migration affects only the minimum number of PPCU tasks, while memory pre-copying technology keeps migration latency below 300ms. For nodes that are chronically overloaded, their weights are automatically reduced, and a scaling recommendation report is generated.

[0086] Furthermore, a multi-layered isolation environment is constructed on the task execution nodes: the hardware layer enables a TEE security zone to restrict memory access; the operating system layer deploys a sandbox to intercept dangerous system calls; and the runtime layer employs the WebAssembly security module to prohibit unauthorized memory access. Each PPCU task runs in an independent container instance, with hard resource isolation achieved through cgroup v2 to prevent side-channel attacks.

[0087] In some embodiments of the present invention, such as Figure 5 As shown, step S105 includes: S501. Based on the allocation credentials, perform cross-validation on the node groups that have completed the same privacy-preserving computing task package to obtain the cross-validation results.

[0088] Node identity verification refers to the digital credential used to verify the legitimate identity of the receiving node. This can be implemented using digital signature technology based on asymmetric encryption algorithms, ensuring the authenticity and non-repudiation of the node's identity through signature verification. Task hash refers to the unique identifier generated for the task content using a hash algorithm, specifically the SHA-256 algorithm, used to quickly detect whether task data has been tampered with during the verification process. The execution time window refers to the start and end time range within which the task is allowed to execute. This can be implemented using timestamps and trusted clock synchronization protocols, used to constrain the task execution cycle to prevent timeouts or delays. Cross-validation involves comparing the processing results of multiple nodes on the same task to detect consistency. This can be implemented using majority voting mechanisms or redundant computation comparison methods, used to detect abnormal nodes or data tampering.

[0089] In a specific embodiment, the verification process is automatically triggered after the computing node completes the PPCU task, first constructing a verification context environment. This environment includes task metadata (such as PPCU ID, computation hash, etc.), node identity credentials (TEEattestation report), and verification strategy configuration (allowed error range, privacy budget, etc.). The verification controller obtains the smart contract corresponding to the task from the blockchain, loads preset verification rules and threshold parameters, and establishes a benchmark standard for subsequent verification steps. The computing node generates an execution proof package in its TEE secure environment, containing three key data parts: first, a hash chain of the complete execution trajectory, recording each intermediate state from input loading to result output; second, zero-knowledge proofs (such as zk-STARK), proving that the computation process conforms to predetermined logic and does not access data without authorization; and third, resource usage audit logs, recording hardware-level indicators such as CPU cycles and memory access patterns. These proof data are signed with the node's private key and divided into multiple fragments stored on different verification nodes. Cross-validation is performed on a group of nodes (usually 3-5) that have completed the same privacy-preserving computing task package based on the allocated credentials, obtaining the cross-validation results. Specifically, an improved PBFT protocol is adopted, which is completed through a three-phase interaction: first, each node broadcasts its own cryptographic commitment to the result; then, partial result fragments are exchanged based on the Merkle tree structure; finally, threshold signatures are executed to reach a consensus on the consistent result. For numerical results, fluctuations within a preset error range are allowed (e.g., ±0.5%); for classification results, the Cohen's Kappa coefficient is required to be ≥0.75.

[0090] S502. Randomly select at least one privacy-preserving computing task package for re-execution and spot check verification to obtain the integrity verification result.

[0091] The re-execution sampling verification involves randomly selecting a portion of tasks and re-executing them in a trusted environment to verify the correctness of the results. This can be achieved through isolated execution in a pre-set sandbox environment, used to dynamically detect potential computational errors or malicious code injection. The verification node verifies the computation process through re-execution sampling: randomly selecting 5% of PPCU tasks and re-executing them in the sandbox environment using the same input. Then, the hardware characteristics of the two executions (such as CPU instruction count and cache hit rate) are compared to detect any abnormal deviations. Simultaneously, the memory access pattern is checked to ensure it meets expectations, preventing data leakage through covert channels, thus obtaining integrity verification results. All performance metrics generated during the verification process are recorded in a time-series database to establish a baseline for node behavior.

[0092] S503. Conduct privacy compliance audits on multiple privacy-preserving computing task packages and obtain the audit results.

[0093] Privacy compliance auditing refers to checking whether data access and computation during task processing comply with privacy protection rules. This can be achieved through a predefined policy library and an automated rule engine to ensure data processing meets legal and regulatory requirements. A dedicated privacy auditing module checks: whether differential privacy budget consumption exceeds the allocated value (ε≤ε_max); whether k-anonymization is correctly implemented (≥k records per group); and whether data lineage is complete and traceable. Sampling verification techniques are used during auditing, probabilistically decrypting sensitive fields to obtain audit results, ensuring audit effectiveness while avoiding large-scale exposure of raw data. Upon detection of violations, a data erasure protocol is automatically triggered, and the relevant nodes are marked as untrusted.

[0094] S504. Verify the logical consistency of multiple privacy-preserving computation task packages based on the directed acyclic graph to obtain the global dependency verification results.

[0095] Logical consistency verification involves checking the correctness of the execution order and result delivery of each subtask based on task dependencies. This can be achieved through topological sorting of a directed acyclic graph (DAG) and a state machine model, preventing overall task failure due to incorrect dependencies. Based on the initial DAG, the logical consistency between the results of each PPCU is verified. All intermediate results are indexed by constructing a Merkle Patricia tree, requiring nodes to provide proof of the input-output connection between adjacent PPCUs. Circular dependencies and deadlock risks are specifically checked to ensure that the final result of distributed computing is equivalent to centralized execution, yielding global dependency verification results. For machine learning tasks, the consistency of gradient update direction is also verified to prevent model divergence caused by malicious nodes.

[0096] S505 generates distributed verification results based on threshold cryptography, cross-validation results, integrity verification results, audit results, and global dependency verification results.

[0097] Threshold cryptography refers to encryption methods that require collaboration among multiple participants to decrypt or generate valid results. Specifically, it can be implemented using the Shamir secret sharing scheme or distributed key generation protocols to prevent single-point tampering or deception in distributed verification. Threshold cryptography requires more than two-thirds of the verification nodes to jointly sign in order to generate the final verification report (i.e., the distributed verification result). The report includes: the verification status of each PPCU, anomaly event records, resource usage statistics, and privacy protection proofs. All intermediate verification data is organized into a Merkle forest structure, with the root hash value anchored to the blockchain, supporting auditing and review at any subsequent verification stage.

[0098] Specifically, after allocating the privacy-preserving computation task packages, the identity of the receiving nodes is first verified to ensure they have the legal authority to execute the tasks. Then, the consistency between the task content obtained by the receiving nodes and the original data is verified through task hash comparison. For multiple nodes in the same task group, a cross-validation mechanism is used to compare their output results; if a difference exceeds a preset threshold, an anomaly alarm is triggered. Simultaneously, a portion of the task packages are randomly selected and re-executed in an independent environment, and the results are compared with the original execution records for completeness. The privacy compliance audit module automatically scans data access logs and encryption operation records during task execution to identify behaviors that violate privacy protection rules. The logical consistency verification module checks whether the input and output of each subtask conform to a preset order based on the task dependency graph, preventing overall task failure due to logical errors. Finally, threshold cryptography is used to encrypt and aggregate the above multi-dimensional verification results; the final distributed verification result is generated only when a preset threshold number of verifiers confirm the result as valid.

[0099] Furthermore, after the receiving node completes its task, its reputation score is updated in real time based on the verification results, using a multi-factor weighted algorithm: task completion rate (30%), result accuracy (40%), response timeliness (20%), and resource consumption rate (10%). The new score is updated via a blockchain smart contract and influences the weighting of subsequent task allocations. For nodes exhibiting abnormal performance, in-depth diagnostics are initiated, including hardware fingerprint comparison and historical behavior pattern analysis, to distinguish between temporary malfunctions and malicious behavior.

[0100] Furthermore, a tiered response mechanism is implemented: for temporary errors (such as network timeouts), a retry mechanism is automatically triggered (up to 3 times); for persistent anomalies (such as continuously deviating results), a backup computing node replacement process is initiated; for security violations (such as evidence of privacy breaches), the relevant nodes are immediately isolated and the regulatory authorities are notified. All recovery operations are recorded in an immutable audit log to ensure accountability.

[0101] Furthermore, a pipelined verification architecture is adopted, allowing the verification and computation phases to partially overlap. Innovative incremental verification technology is used, re-verifying only the changed parts, reducing typical verification latency by 40%. For high-frequency repetitive tasks, a verification result cache pool is established, reusing historical verification conclusions through similarity matching. Simultaneously, FPGA accelerators are deployed to specifically optimize computationally intensive operations such as zero-knowledge proof verification.

[0102] In some embodiments of the present invention, such as Figure 6 As shown, step S106 includes: S601. Perform integrity verification on all privacy-preserving computing task packages that participate in the aggregation of all distributed verification results to obtain the verification results.

[0103] Integrity verification refers to verifying that all privacy-preserving computation task packages have not been tampered with or lost during transmission and processing. This can be achieved using hash chain verification or Merkle tree verification mechanisms to ensure the authenticity of the data source and the reliability of the transmission process. First, integrity verification is performed on all PPCU computation results participating in the aggregation. This is done by verifying the Merkle proof chain to ensure the data has not been tampered with, thus obtaining the verification result. Each result package must contain signature confirmations from at least three independent nodes, and the timestamp must be checked to ensure it is within a valid time window (typically set to 1.5 times the computation timeout). For numerical data, outliers exceeding 3 times the median absolute deviation (MAD) are automatically filtered out; for categorical data, it is checked whether it conforms to a predefined set of enumerated values.

[0104] S602. After cleaning the data of multiple privacy-preserving computing task packages according to the verification results, the multiple privacy-preserving computing task packages are aggregated in three levels to obtain the final aggregation result; the levels in the three-level aggregation include node level, computing cluster level and aggregation level.

[0105] Data cleaning refers to filtering or correcting task packages containing redundancy, errors, or conflicts. This can be achieved through rule engines or anomaly detection algorithms to improve the accuracy of subsequent aggregation operations. The data cleaning process records a complete audit trail, including feature analysis and processing basis for the removed data. Three-level aggregation refers to progressively merging calculation results at different levels. Specifically, it can employ a layered processing approach: node-level local aggregation, computation cluster-level regional aggregation, and global-level cross-domain aggregation, to balance computational load and reduce network communication overhead. Specifically, it executes three-level aggregation from bottom to top, following the reverse process of the initial task decomposition.

[0106] S603. Perform a four-dimensional evaluation on the final aggregation result to obtain the encrypted computing radiation scheduling result; the four-dimensional evaluation includes logical consistency, statistical rationality, computational correctness and privacy compliance.

[0107] The four-dimensional evaluation refers to the quantitative analysis of the aggregation results from four dimensions: security, efficiency, resource utilization, and task completion. This can be achieved through a pre-set weighted model or a dynamic scoring algorithm, used to comprehensively judge the optimization degree of the scheduling results. The final aggregation result undergoes a four-dimensional evaluation: logical consistency (checking compliance with domain knowledge rules), statistical reasonableness (verifying distribution characteristics), computational correctness (sampling and recalculating for verification), and privacy compliance (ε-consumption audit). The evaluation process generates a detailed quality scorecard, including scores for each dimension and improvement suggestions. For critical task scenarios, multiple candidate result versions (such as aggregation results under different privacy budgets) are retained to obtain encrypted computation radiation scheduling results, which end users can choose according to their actual needs.

[0108] Specifically, the integrity verification phase identifies anomalous data units by comparing the hash value of the task package with distributed verification records. Subsequently, the data cleaning module removes invalid task packages based on the verification results and repairs some recoverable data errors. In the three-level aggregation process, each receiving node first performs preliminary aggregation on its local task packages, generating node-level intermediate results. Then, at the computing cluster level, the intermediate results from multiple nodes are aggregated again to form regional summary data. Finally, at the global level, all regional data are integrated, and the final aggregation is completed through an encrypted channel. After aggregation, the four-dimensional evaluation module calculates the security vulnerability detection rate, average task processing latency, resource utilization, and task completion rate, generating a multi-dimensional scoring matrix as the basis for optimizing the scheduling results.

[0109] Through the above technical solutions, this application solves the problem of balancing privacy protection and computational efficiency in existing technologies, and realizes secure aggregation and multi-dimensional evaluation of cross-level data. Integrity verification and data cleaning ensure the reliability of input data, the three-level aggregation mechanism reduces network transmission pressure, and the four-dimensional evaluation model provides a quantitative basis for scheduling strategy optimization, thereby improving resource utilization and task execution success rate in complex heterogeneous environments.

[0110] In some embodiments of the present invention, such as Figure 7 As shown, step S602 includes: S701. Perform three-level aggregation on multiple privacy-preserving computing task packages to obtain multi-level result alignment results.

[0111] In a specific implementation, firstly, at the node level, multiple replicas of the same PPCU are merged using a robust weighted average algorithm (the weights of each node are dynamically adjusted based on its historical accuracy). Then, at the cluster level, the outputs of related PPCU groups are fused, and feature embedding spatial alignment technology is used to address the dimensionality shift issue caused by distributed computing. Finally, at the global level, all intermediate results are integrated, and a recursive merging algorithm is used to process complex data structures, maintaining the complete semantics of nested JSON / XML formats, resulting in multi-level result alignment. Each aggregation level generates a corresponding quality assessment report, including confidence intervals, analysis of variance, and other indicators.

[0112] S702. Dynamically inject differential privacy protection into the multi-level result alignment results to obtain privacy-enhanced processing results.

[0113] Dynamic injection of differential privacy protection refers to a technique that adjusts the noise injection intensity in real time according to a preset privacy budget during data aggregation. Specifically, it can be implemented using a Laplace mechanism combined with a dynamic budget allocation algorithm, enhancing privacy protection while ensuring data availability. During aggregation, dynamic injection of differential privacy protection involves adding calibrated Laplace noise (noise scale = Δf / ε, where Δf is the global sensitivity) to numerical results and using an exponential mechanism for probability perturbation of categorical data. Privacy budget consumption is monitored in real time, and degradation processing (such as replacing precise values ​​with ranges) is automatically triggered when ε exceeds 80%. Simultaneously, k-anonymization post-processing is implemented to ensure that each aggregation result corresponds to at least k original inputs (default k=50). Rare feature combinations are eliminated through generalization, suppression, and other techniques to obtain privacy-enhanced results. All privacy operations are recorded in a dedicated privacy ledger for later auditing.

[0114] S703. The privacy enhancement processing results are encrypted and aggregated through a secure multi-party computation protocol to obtain the multi-party computation aggregation result.

[0115] Secure multi-party computation protocol encryption aggregation refers to a technique that uses cryptographic methods to achieve collaborative data computation among multiple parties. Specifically, it can be implemented using a Shamir threshold scheme based on secret sharing combined with homomorphic encryption algorithms, ensuring that participating parties cannot obtain the original data from other nodes. For particularly sensitive statistics (such as sums and means), a secure multi-party computation protocol is used: data is distributed to multiple participating parties in an encrypted state through secret sharing technology. The Shamir threshold scheme ensures that fragments exceeding a threshold must be collected to reconstruct the result, obtaining the multi-party computation aggregation result. For complex calculations such as matrix operations, aggregation operations are performed directly on the ciphertext based on homomorphic encryption characteristics. The optimal scheme is automatically selected based on the data type—additive homomorphism (Paillier) is used for numerical accumulation, and fully homomorphic (CKKS) is used for polynomial calculations.

[0116] S704. Perform anomaly detection and conflict resolution on the multi-party computation aggregation results to obtain the final aggregation result.

[0117] Anomaly detection and conflict resolution refer to the process of verifying data consistency and correcting errors in the aggregation results. Specifically, this can be achieved using an anomaly detection model based on the Isolation Forest algorithm combined with a majority voting mechanism to eliminate erroneous data caused by node failures or malicious attacks. A machine learning-based anomaly detection system is deployed to analyze data distribution changes in real time during the aggregation process. When a sudden change in statistical characteristics is detected (such as a sudden increase of 30% in standard deviation), root cause analysis is automatically triggered to check for possible factors such as changes in participating node reputation and network latency fluctuations. For result conflicts, a Trusted Arbitration Mechanism based on TEE is used, running a predefined conflict resolution rule base within a secure enclave. If necessary, a manual review process is introduced to obtain the final aggregation result. All anomaly events generate reports containing complete evidence chains for subsequent optimization reference.

[0118] Specifically, after completing the distributed verification of the privacy-preserving computation task package, a three-level aggregation process is first executed in the order of node level, computation cluster level, and aggregation level. Node-level aggregation initially integrates the task package results within the same computation node; computation cluster-level aggregation merges data from related nodes according to logical relationships; and the aggregation level completes the unified processing of global data. After each level of processing is completed, cross-level data alignment is ensured through timestamps and version identifiers. Subsequently, differential privacy noise is dynamically injected into the intermediate aggregation results, with the noise intensity dynamically adjusted according to data sensitivity. After privacy enhancement, a secure multi-party computation protocol is used to encrypt and aggregate the intermediate results scattered across the participants, ensuring that the data remains in an encrypted state throughout the aggregation process. Finally, a rule engine detects abnormal data patterns in the aggregation results, such as out-of-bounds values ​​or broken logical dependencies, and automatically corrects conflicting data based on preset strategies, outputting the final aggregation result.

[0119] Through the above technical solutions, this application can solve the data inconsistency problem caused by single-layer aggregation in traditional heterogeneous resource scheduling, and improve data integration efficiency through a hierarchical aggregation structure. The dynamic differential privacy injection mechanism can adjust the privacy protection strength according to task requirements, avoiding excessive privacy loss or insufficient protection caused by fixed noise parameters. The combination of secure multi-party computation protocol and anomaly detection can achieve highly reliable global aggregation in an encrypted environment, effectively preventing malicious nodes from tampering with the aggregation results, thereby improving the accuracy and reliability of encrypted computation-based scheduling results.

[0120] Furthermore, after aggregation is complete, a comprehensive resource cleanup can be performed: securely erasing all intermediate computation data (compliant with NIST SP 800-88 standards), releasing distributed computing node resources, and updating the reputation scores of participating nodes. Simultaneously, a task execution report is generated, including resource usage statistics, performance metrics, and a summary of anomalies. All cleanup operations are recorded on the blockchain ledger, ensuring auditability throughout the entire lifecycle.

[0121] Furthermore, various metrics during the aggregation process (such as time consumption at each stage and changes in data quality) are collected and input into the reinforcement learning model to optimize parameters. Performance bottlenecks (such as aggregation latency caused by network delays) are automatically identified, and suggestions for topology adjustments or algorithm parameter optimization schemes are generated. For repetitive tasks, an aggregation pattern knowledge base is established, and case matching accelerates the execution of subsequent similar tasks.

[0122] like Figure 8 As shown, the multi-terminal interaction process in this embodiment of the invention is as follows: The optimization terminal submits the encrypted computation result. After receiving the result, the radiation center decomposes the task to obtain a set of PPCUs. Then, each PPCU is transmitted to the corresponding computing node through the radiation allocation and verification network. For example, PPCU-1 is sent to computing node 1, PPCU-2 is sent to computing node 2, and PPCU-N is sent to computing node N. After transmission, the computing node submits the intermediate result hash to the verification network. The verification network (i.e., the dynamic topology network) returns the corresponding cross-validation result to the corresponding computing node. Then, the computing node returns the verified result to the radiation center. The radiation center aggregates all the verification results to obtain the final aggregated result and returns the encrypted computation result to the user terminal based on the final aggregated result.

[0123] This invention proposes a quantum-classical hybrid scheduling model that supports seamless task migration between QPU and CPU / GPU. End-to-end privacy protection is achieved based on quantum homomorphic encryption (QHE) to resist quantum attacks. A quantum decoherence adaptive compensation algorithm is introduced to improve the stability of quantum resources.

[0124] To better implement the heterogeneous computing resource capacity radiation scheduling method in the embodiments of the present invention, correspondingly, the embodiments of the present invention also provide a heterogeneous computing resource capacity radiation scheduling device, such as... Figure 9 As shown, the heterogeneous computing resource capacity radiation scheduling device 900 includes: The task receiving module 901 is used to acquire the original encrypted computing task and confirm the computing capabilities of multiple receiving nodes in the radiation center. The task package decomposition module 902 is used to decompose the original encrypted computing task into multiple privacy-preserving computing task packages; The node matching module 903 is used to confirm the node matching result based on the computing power of multiple receiving nodes and the Pareto optimal solution of multiple privacy-preserving computing task packages. The task package allocation module 904 is used to allocate multiple privacy-preserving computing task packages to the corresponding receiving nodes according to the node matching results, and obtain the allocation credential for each receiving node. The allocation verification module 905 is used to perform distributed verification on the allocation credential and obtain the distributed verification result. The multimodal aggregation module 906 is used to perform multimodal aggregation on all distributed verification results to obtain the encrypted computing radiation scheduling result.

[0125] The heterogeneous computing resource capacity radiation scheduling device 900 provided in the above embodiments can realize the technical solutions described in the above embodiments of the heterogeneous computing resource capacity radiation scheduling method. The specific implementation principles of each module or unit can be found in the corresponding content in the above embodiments of the heterogeneous computing resource capacity radiation scheduling method, which will not be repeated here.

[0126] The heterogeneous computing resource capacity radiation scheduling method and apparatus provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for heterogeneous computing resource capacity radiation scheduling, characterized in that, include: Obtain the original encrypted computation task and confirm the computing capabilities of multiple receiving nodes in the radiation center; The original encrypted computing task is broken down into multiple privacy-preserving computing task packages; Based on the computing power of the plurality of receiving nodes and the Pareto optimal solution of the plurality of privacy-preserving computing task packages, the node matching result is confirmed; Based on the node matching results, the multiple privacy-preserving computing task packages are allocated to the corresponding receiving nodes, and the allocation credential for each receiving node is obtained; The allocation credential is subjected to distributed verification to obtain the distributed verification result; Multimodal aggregation of all distributed verification results yields encrypted computation radiation scheduling results.

2. The heterogeneous computing resource capacity radiation scheduling method according to claim 1, characterized in that, The verification of the computing capabilities of multiple receiving nodes in the radiation center includes: Acquire basic data and historical reliability records of the multiple receiving nodes at the radiation center; the basic data includes data in the dimensions of computation, storage, security, and network. A multidimensional capability profile for each receiving node is constructed based on the data from the computational dimension, the storage dimension, the security dimension, and the network dimension. The corresponding computing capabilities are obtained based on the multidimensional capability profile and historical reliability record of each receiving node.

3. The heterogeneous computing resource capacity radiation scheduling method according to claim 1, characterized in that, The metadata portion of the original encrypted computation task includes structured and unstructured data; the step of decomposing the original encrypted computation task into multiple privacy-preserving computation task packages includes: After decrypting the metadata part of the original encrypted computing task, the fields in the structured data are encrypted according to the preset sensitivity of the fields to obtain structured encrypted data. Then, the structured encrypted data is subjected to multi-dimensional feature clustering to obtain multiple feature information. The unstructured data is detected and segmented according to a preset target detection model to obtain segmented region information; The multiple feature information and the block region information are stored in multiple segments of the radiation center to obtain a mapping table for each segment; The original encryption computation task is converted into a directed acyclic graph, and the directed acyclic graph is decomposed to obtain the decomposition result; Based on the decomposition results and the mapping table, multiple initial privacy-preserving computation task packages are generated; After generating a unique privacy credential based on the data in the fragments of each initial privacy-preserving computation task package, four checks are performed on each initial privacy-preserving computation task package based on the unique privacy credential to obtain the check results; the four checks include data integrity, logical completeness, privacy compliance, and performance prediction. The initial privacy-preserving computation task package that passes the inspection is marked as ready, resulting in multiple privacy-preserving computation task packages marked as ready.

4. The heterogeneous computing resource capacity radiation scheduling method according to claim 1, characterized in that, After decomposing the original encryption computation task into multiple privacy-preserving computation task packages, the method further includes: After the original encrypted computing task is broken down by bandwidth, if the bandwidth decreases by a preset percentage, the fragments in the multiple privacy-preserving computing task packages are split and recombined to obtain multiple privacy-preserving computing task packages with smaller capacity.

5. The heterogeneous computing resource capacity radiation scheduling method according to claim 3, characterized in that, The step of confirming the node matching result based on the computing power of the plurality of receiving nodes and the Pareto optimal solution of the plurality of privacy-preserving computing task packages includes: The multiple receiving nodes are evaluated based on the evaluation indicators and the computing power to obtain the evaluation results; The execution order constraints of the multiple privacy-preserving computation task packages are determined based on the dependencies in the directed acyclic graph. Based on the NSGA-II algorithm, the Pareto optimal solution for each privacy-preserving computation task package is determined among the four checks of the multiple privacy-preserving computation task packages in terms of privacy protection strength, computational efficiency, and cost. The node matching result is confirmed based on the execution order constraint, the Pareto optimal solution, and the evaluation result.

6. The heterogeneous computing resource capacity radiation scheduling method according to claim 4, characterized in that, The step of allocating the multiple privacy-preserving computation task packages to corresponding receiving nodes based on the node matching results, and obtaining the allocation credential for each receiving node, includes: Set up encrypted transmission channels and fault tolerance mechanisms; Each privacy-preserving computing task package is transmitted to the corresponding receiving node through the dynamic topology network, the encrypted transmission channel, and the node matching result. The transmission process is monitored according to the fault tolerance mechanism to obtain the monitoring result. When the monitoring result is abnormal, the corresponding recovery process is initiated according to the fault type in the monitoring result, so that each privacy protection computing task package is transmitted to the corresponding receiving node; When the receiving node finishes receiving, it obtains the allocation credential for each receiving node.

7. The heterogeneous computing resource capacity radiation scheduling method according to claim 3, characterized in that, The distributed verification of the allocation credential to obtain the distributed verification result includes: Based on the allocation credentials, cross-validation is performed on the node groups that have completed the same privacy-preserving computation task package to obtain the cross-validation results; Randomly select at least one privacy-preserving computation task package for re-execution and spot check verification to obtain the integrity verification result; A privacy compliance audit was performed on the multiple privacy-preserving computing task packages, and the audit results were obtained. The logical consistency of the multiple privacy-preserving computation task packages is verified based on the directed acyclic graph to obtain the global dependency verification results. A distributed verification result is generated based on threshold cryptography, the cross-validation result, the integrity verification result, the audit result, and the global dependency verification result.

8. The heterogeneous computing resource capacity radiation scheduling method according to claim 1, characterized in that, The process of multimodal aggregation of all distributed verification results to obtain encrypted computation radiation scheduling results includes: Integrity verification is performed on all privacy-preserving computation task packages that participate in the aggregation of all distributed verification results to obtain the verification result; After cleaning the data of the multiple privacy-preserving computing task packages based on the verification results, the multiple privacy-preserving computing task packages are then aggregated in three levels to obtain the final aggregation result; the three-level aggregation includes node level, computing cluster level and aggregation level. The final aggregation result is evaluated in four dimensions to obtain the encrypted computation radiation scheduling result; the four-dimensional evaluation includes logical consistency, statistical rationality, computational correctness and privacy compliance.

9. The heterogeneous computing resource capacity radiation scheduling method according to claim 8, characterized in that, The three-level aggregation of the multiple privacy-preserving computation task packages yields the final aggregation result, including: The multiple privacy-preserving computation task packages are aggregated at three levels to obtain multi-level result alignment results; Differential privacy protection is dynamically injected into the alignment results of the multi-level results to obtain privacy-enhancing processing results; The privacy-enhancing processing results are encrypted and aggregated using a secure multi-party computation protocol to obtain the multi-party computation aggregation result; The multi-party computation aggregation results are subjected to anomaly detection and conflict resolution to obtain the final aggregation result.

10. A heterogeneous computing resource capacity radiation scheduling device, characterized in that, include: The task receiving module is used to acquire the original encrypted computing task and confirm the computing capabilities of multiple receiving nodes in the radiation center. The task package decomposition module is used to decompose the original encrypted computing task into multiple privacy-preserving computing task packages; The node matching module is used to confirm the node matching result based on the computing power of the plurality of receiving nodes and the Pareto optimal solution of the plurality of privacy-preserving computing task packages; The task package allocation module is used to allocate the multiple privacy-preserving computing task packages to the corresponding receiving nodes according to the node matching results, and obtain the allocation credential for each receiving node; The allocation verification module is used to perform distributed verification on the allocation credential and obtain the distributed verification result. The multimodal aggregation module is used to aggregate all distributed verification results in a multimodal manner to obtain the encrypted computing radiation scheduling result.

Images