Data processing method and system based on fusion of multi-party secure computation and federated learning

Abstract

This application discloses a data processing method and system based on the fusion of multi-party secure computation and federated learning. The method includes: in response to a data fusion request, obtaining the gradient update features of the business dataset and the current federated learning strategy; generating a multi-party secure computation protocol based on the gradient update features, including node interaction order rules, privacy-preserving computation logic rules, and protocol update frequency rules; using the gradient data as protocol input, performing privacy-preserving fusion of the gradient data and the business dataset through lightweight encrypted interaction to obtain a fused encrypted dataset; extracting computation requirement sub-features and encryption requirement sub-features, and simultaneously optimizing the federated learning strategy and the multi-party secure computation protocol based on their correlation mapping, so that the optimized gradient computation process and the encryption processing process work together, realizing a deep fusion of federated learning and multi-party secure computation, ensuring data privacy while improving data processing efficiency.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a data processing method and system based on the fusion of multi-party secure computation and federated learning. Background Technology

[0002] Against the backdrop of accelerated digital transformation, the demand for multi-source data integration in target business scenarios (such as fintech risk control, R&D data collaboration, and enterprise resource optimization) is becoming increasingly urgent. Federated learning, as a distributed machine learning technology, can achieve collaborative training of multi-node models without sharing the original data. Its core lies in updating the global model by uploading gradient data from each participating node, effectively solving the data silo problem. Multi-party secure computation, on the other hand, uses cryptographic technology to achieve joint computation of multi-node data, ensuring the privacy and security of data during the computation process. The combination of these two technologies has become a key technological direction for resolving the contradiction between data privacy and data value mining.

[0003] In existing technologies, the formulation of federated learning strategies is often independent of multi-party secure computation protocols, making it difficult to coordinate the gradient calculation process of federated learning with the encryption process of multi-party secure computation. At the same time, traditional data fusion methods often use heavyweight encryption algorithms, which not only increases computational overhead but also reduces the efficiency of data fusion. Furthermore, the optimization of federated learning strategies focuses more on improving model accuracy, and lacks attention to the coordinated optimization with multi-party secure computation protocols, making it difficult to meet the needs of target business scenarios for secure and efficient data processing.

[0004] Therefore, in the process of data processing based on the integration of federated learning strategies and multi-party secure computation protocols, it is difficult to balance data privacy and processing efficiency. Summary of the Invention

[0005] This application provides a data processing method and system based on the fusion of multi-party secure computation and federated learning, which can achieve deep integration of federated learning and multi-party secure computation, and improve data processing efficiency while protecting data privacy.

[0006] This application provides a data processing method based on the fusion of multi-party secure computation and federated learning, applied to a data processing system. The method includes: In response to a data fusion request for a target business scenario, the business dataset of the target business scenario and the gradient update features of the current federated learning strategy are obtained. A multi-party secure computation protocol is generated based on the gradient update features. The multi-party secure computation protocol includes rules for the node interaction order during protocol execution, rules for privacy-preserving computation logic, and rules for protocol update frequency. The gradient data of the current federated learning strategy is used as the input of the multi-party secure computation protocol. The gradient data and the business dataset are fused through a lightweight encrypted interaction to obtain a fused encrypted dataset. Extract the computational requirement sub-features corresponding to the gradient training features of the current federated learning policy during the policy training phase, as well as the encryption requirement sub-features of the fused encrypted dataset; wherein, the computational requirement sub-features include information on the number of computation nodes for gradient aggregation, information on the length of the computation path for gradient propagation, and information on the frequency of gradient computation, and the encryption requirement sub-features include information on the range of decryption nodes for encrypted gradient data, information on the length of the decryption path for encrypted business data, and information on the encryption update frequency. Based on the association mapping between the computational requirement sub-features and the encryption requirement sub-features, the current federated learning strategy and the multi-party secure computation protocol are synchronously optimized, so that the gradient calculation process of the optimized federated learning strategy and the encryption processing process of the optimized multi-party secure computation protocol are coordinated.

[0007] One embodiment of this application provides a data processing system, including: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement any of the data processing methods based on the fusion of multi-party secure computation and federated learning.

[0008] One embodiment of this application provides a readable storage medium on which a program or instruction is stored. When the program or instruction is executed by a processor, it implements the steps of the data processing method based on the fusion of multi-party secure computation and federated learning.

[0009] This application's embodiments achieve deep integration of federated learning and multi-party secure computation through a full-link collaborative logic involving gradient update feature-driven multi-party secure computation protocol generation, lightweight encryption fusion, requirement sub-feature extraction, and dual-strategy synchronous optimization. First, an adapted multi-party secure computation protocol is generated based on the gradient update features of the current federated learning strategy, ensuring that the protocol's node interaction order, privacy protection logic, and update frequency accurately match the gradient characteristics of federated learning. Second, lightweight encryption interaction completes the privacy-preserving fusion of gradient data and business datasets, effectively integrating data value while ensuring data privacy. Third, by extracting computation requirement sub-features and encryption requirement sub-features, an association mapping relationship is constructed, providing a quantitative basis for dual-strategy synchronous optimization. Finally, based on this mapping relationship, the federated learning strategy and the multi-party secure computation protocol are synchronously optimized, enabling the gradient computation process of the optimized federated learning strategy and the encryption processing process of the optimized multi-party secure computation protocol to collaborate.

[0010] This design breaks through the bottlenecks of efficiency loss and high privacy leakage risk caused by the independent operation of traditional federated learning and multi-party secure computation. While ensuring data privacy, it improves the efficiency of data fusion and the accuracy of model optimization, thereby achieving secure and efficient data processing in the target business scenario. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating a data processing method based on the fusion of multi-party secure computation and federated learning, provided as an embodiment of this application.

[0013] Figure 2 This is a schematic diagram of the basic structure of a data processing system provided in an embodiment of this application.

[0014] Figure 3 This is a functional block diagram of a data processing device provided in an embodiment of this application. Detailed Implementation

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0016] Please see Figure 1 , Figure 1 This is a flowchart of a data processing method based on the fusion of multi-party secure computation and federated learning provided in an embodiment of this application. The method can be executed by a data processing system or jointly executed by a data processing system and a server. The method may include steps 110-150.

[0017] This application provides a data processing method based on the fusion of multi-party secure computation and federated learning, applied to a data processing scenario in a cloud-based collaborative office environment where multi-party secure computation and federated learning are integrated. The executing entity is a data processing system. The following describes each step in detail with reference to this scenario.

[0018] Step 110: In response to the data fusion request for the target business scenario, obtain the business dataset of the target business scenario and the gradient update features of the current federated learning strategy.

[0019] In this embodiment, the target business scenario is a data processing scenario that integrates multi-party secure computation and federated learning in a cloud-based collaborative office environment. The data fusion request is initiated by a participating node of the cloud office platform, and the request content is to merge the office data of each participating node to optimize the collaborative office model. After receiving the request, the data processing system first obtains the business dataset, which includes document editing operation data of participating node A, meeting arrangement data of participating node B, task allocation data of participating node C, etc. All of the above data are non-sensitive office data stored locally by each node. Next, it obtains the gradient update features of the current federated learning strategy. The current federated learning strategy is a federated learning strategy used to optimize the collaborative office model. Its gradient update features include the direction of gradient update, the magnitude of gradient update, and the nodes involved in the gradient update. For example, the gradient update direction is to improve the collaboration between document editing and meeting arrangement, the gradient update magnitude is medium, and the gradient update involves nodes A, B, and C.

[0020] Step 120: Generate a multi-party secure computation protocol based on the gradient update features. The multi-party secure computation protocol includes rules for the node interaction order of protocol execution, rules for privacy-preserving computation logic, and rules for protocol update frequency.

[0021] In this embodiment, the data processing system generates a multi-party secure computation protocol based on gradient update characteristics. First, the node interaction order rule is determined. Since the gradient update involves nodes A, B, and C, and the document editing data of node A is highly correlated with the meeting arrangement data of node B, the node interaction order rule is set so that node A interacts with node B first, and then with node C. Next, privacy-preserving computation logic rules are determined, employing homomorphic encryption technology to encrypt the interaction data, ensuring that the original information is not leaked during computation. Finally, the protocol update frequency rule is determined. Based on the moderate gradient update amplitude, the protocol update frequency is set to once every three days.

[0022] Step 130: Use the gradient data of the current federated learning strategy as input to the multi-party secure computation protocol, and perform privacy-preserving fusion processing on the gradient data and the business dataset through lightweight encrypted interaction to obtain a fused encrypted dataset.

[0023] In this embodiment, the gradient data of the current federated learning strategy are gradient parameters used to optimize the office collaboration model. These gradient parameters include document editing feature weights, meeting scheduling feature weights, and task allocation feature weights. The data processing system inputs this gradient data into a multi-party secure computation protocol for lightweight encrypted interaction. First, the gradient data is encrypted using a lightweight symmetric encryption algorithm, with the key generated through negotiation among participating nodes. Then, the business dataset is encrypted using the same symmetric encryption algorithm. Next, the encrypted gradient data and the encrypted business dataset are fused. The fusion method involves weighted fusion of each feature weight in the gradient data with the corresponding feature data in the business dataset. For example, the document editing feature weight is weighted and fused with the document editing operation data of node A; the meeting scheduling feature weight is weighted and fused with the meeting scheduling data of node B; and the task allocation feature weight is weighted and fused with the task allocation data of node C. The result is a fused encrypted dataset.

[0024] Step 131: Perform gradient feature association processing on the gradient propagation path in the gradient update features and the feature values ​​of the gradient data of the current federated learning policy to generate a gradient feature association matrix.

[0025] In this embodiment, the gradient propagation path in the gradient update features is node A → node B → node C. The feature values ​​of the gradient data of the current federated learning strategy include document editing feature value W1, meeting scheduling feature value W2, and task allocation feature value W3. The data processing system associates the gradient propagation path with the above feature values ​​to generate a gradient feature association matrix. The rows of the matrix represent nodes on the gradient propagation path, the columns represent feature values, and the matrix elements represent the degree of association between the node and the feature value. For example, the degree of association between node A and W1 is 0.8, the degree of association between node B and W2 is 0.9, and the degree of association between node C and W3 is 0.7.

[0026] Step 132: Perform business feature association processing on the business type information and feature values ​​corresponding to the business dataset of the target business scenario to generate a business feature association matrix.

[0027] In this embodiment, the business type information corresponding to the business dataset includes document editing business, meeting scheduling business, and task allocation business. The feature values ​​include document editing frequency feature value F1, meeting scheduling duration feature value F2, and task allocation completion rate feature value F3. The data processing system performs association processing on the business type information and feature values ​​to generate a business feature association matrix. The rows of the matrix represent business types, the columns represent feature values, and the matrix elements represent the degree of association between the business type and the feature value. For example, the degree of association between document editing business and F1 is 0.9, the degree of association between meeting scheduling business and F2 is 0.8, and the degree of association between task allocation business and F3 is 0.75.

[0028] Step 133: Based on the association results of the gradient feature association matrix and the business feature association matrix, select a lightweight encryption method for the interaction: if the feature overlap in the association results is higher than the preset overlap, then use a symmetric encryption method to encrypt the gradient data and the business dataset; if the feature overlap is lower than the preset overlap, then use an asymmetric encryption method to encrypt the gradient data and the business dataset.

[0029] In this embodiment, the data processing system first calculates the correlation results of the gradient feature correlation matrix and the business feature correlation matrix, that is, it calculates the sum of the products of corresponding elements in the two matrices to obtain the feature overlap degree. The preset overlap degree is 0.7. If the calculated feature overlap degree is 0.8, which is higher than the preset overlap degree, a symmetric encryption method is used, such as the AES-128 symmetric encryption algorithm; if the calculated feature overlap degree is 0.6, which is lower than the preset overlap degree, an asymmetric encryption method is used, such as the RSA asymmetric encryption algorithm.

[0030] Step 134: Aggregate the feature values ​​of the encrypted gradient data with the feature values ​​of the encrypted business data to generate aggregated encrypted features, and generate a fused encrypted dataset based on the aggregated encrypted features; wherein, the feature association process and the encryption method selection process are carried out simultaneously, and each adjustment of the association result will affect the encryption method selection logic.

[0031] In this embodiment, the data processing system aggregates encrypted gradient data feature values ​​(such as encrypted W1, W2, and W3) with encrypted business data feature values ​​(such as encrypted F1, F2, and F3). The aggregation method involves concatenating the corresponding feature values; for example, concatenating encrypted W1 with encrypted F1 yields aggregated encrypted feature W1F1, and similarly yields W2F2 and W3F3. Then, a fused encrypted dataset is generated based on these aggregated encrypted features. The fused encrypted dataset is a data set containing all aggregated encrypted features. Simultaneously, the feature association processing and encryption method selection processes are performed concurrently. For example, if the element values ​​in the gradient feature association matrix are adjusted during feature association processing, causing a change in feature overlap, the encryption method selection process will immediately reselect the encryption method based on the new feature overlap.

[0032] Step 140: Extract the computational requirement sub-features corresponding to the gradient training features of the current federated learning policy during the policy training phase, and the encryption requirement sub-features of the fused encrypted dataset; wherein, the computational requirement sub-features include information on the number of computation nodes for gradient aggregation, the length of the computation path for gradient propagation, and the frequency of gradient computation, and the encryption requirement sub-features include information on the range of decryption nodes for encrypted gradient data, the length of the decryption path for encrypted business data, and the encryption update frequency.

[0033] In this embodiment, the data processing system first extracts the computational requirement sub-features corresponding to the gradient training features of the current federated learning policy during the policy training phase. Gradient training features include the number of samples for gradient training, the number of iterations for gradient training, and the loss function value for gradient training. The computational requirement sub-features include the number of computation nodes for gradient aggregation (e.g., 3 nodes), the number of nodes traversed for gradient propagation (e.g., 2 nodes), and the frequency of gradient computation (e.g., once per hour). Next, the system extracts the encryption requirement sub-features of the fused encrypted dataset. The encryption requirement sub-features include the decryption node range information for encrypted gradient data (e.g., nodes A, B, and C), the decryption path length information for encrypted business data (e.g., 1), and the encryption update frequency information (e.g., once every two days).

[0034] Step 141: Extract gradient aggregation features, gradient propagation features, and gradient frequency features from the gradient training features of the current federated learning strategy.

[0035] In this embodiment, the data processing system extracts gradient aggregation features from the gradient training features. These features include the gradient aggregation method and the weights of the gradient aggregation. It also extracts gradient propagation features, including the direction and speed of gradient propagation. Furthermore, it extracts gradient frequency features, including the time interval between gradient calculations and the time interval between gradient updates. For example, the gradient aggregation feature uses a weighted average method for gradient aggregation, with weights of 0.4 for node A, 0.3 for node B, and 0.3 for node C. The gradient propagation feature shows that the gradient is propagated from node A to node B, and then to node C, at a speed of 100 MB per second. The gradient frequency feature shows that the gradient calculation time interval is once per hour, and the gradient update time interval is once every two hours.

[0036] Step 142: Perform statistical processing on the gradient aggregation features to generate information on the number of computation nodes for gradient aggregation; perform sequence processing on the gradient propagation features to generate information on the length of the computation path for gradient propagation; perform periodic processing on the gradient frequency features to generate information on the frequency of gradient computation.

[0037] In this embodiment, the data processing system performs statistical processing on gradient aggregation features, counting the number of nodes involved in the gradient aggregation calculation to generate information on the number of computational nodes for gradient aggregation, for example, 3 nodes. It also performs sequence processing on gradient propagation features, counting the number of nodes the gradient must traverse from the starting node to the ending node to generate information on the computational path length for gradient propagation, for example, 2 nodes. Finally, it performs periodic processing on gradient frequency features, counting the time interval between gradient calculations to generate information on the frequency of gradient calculations, for example, once per hour.

[0038] Step 1421: Select feature values ​​related to gradient aggregation from the gradient training features of the current federated learning strategy; perform statistical analysis on the selected feature values, calculate the distribution of the feature values, and determine the number of computing nodes for gradient aggregation based on the distribution of the feature values; if the feature value distribution in the statistical analysis results belongs to the first distribution label, then the number of computing nodes is the number of nodes in a concentrated distribution; if the feature value distribution belongs to the second distribution label, then the number of computing nodes is the number of nodes in a dispersed distribution; dynamically adjust the generated number of computing nodes, and adjust the size of the statistical window for the next time according to the adjustment of the previous statistical results; repeat the statistical analysis and dynamic adjustment steps until the number of computing nodes meets the preset requirements; wherein, the statistical processing process and the dynamic adjustment process are performed synchronously, and each adjustment of the statistical results affects the size of the statistical window for the next time.

[0039] In this embodiment, the data processing system first filters out feature values ​​related to gradient aggregation from the gradient training features, such as gradient aggregation weights and the number of gradient aggregation samples. Then, it performs statistical analysis on these feature values, calculating their distribution, such as the mean, variance, and standard deviation. Based on the distribution of the feature values, it determines the number of computation nodes for gradient aggregation. If the feature value distribution belongs to the first distribution label (e.g., normal distribution), the number of computation nodes is the number of nodes in a concentrated distribution, for example, 3 nodes. If the feature value distribution belongs to the second distribution label (e.g., uniform distribution), the number of computation nodes is the number of nodes in a dispersed distribution, for example, 5 nodes. Next, the generated number of computation nodes is dynamically adjusted. For example, if the number of computation nodes is 3 after the previous statistical adjustment, and if it is found that the number of computation nodes does not match the actual needs, the size of the statistical window is adjusted for the next iteration, for example, from 10 to 20. This statistical analysis and dynamic adjustment process is repeated until the number of computation nodes meets the preset requirements, which are that the number of computation nodes can satisfy the computational needs of gradient aggregation. At the same time, the statistical processing and dynamic adjustment processes are carried out simultaneously. For example, if the size of the statistical window is adjusted during the statistical processing, the dynamic adjustment process will immediately re-perform the statistical analysis based on the new statistical window size.

[0040] Step 143: Combine the computing node quantity information, computing path length information, and computing frequency information into computing requirement sub-features.

[0041] In this embodiment of the application, the data processing system combines the information on the number of computing nodes (e.g., 3 nodes), the information on the length of the computing path (e.g., 2 nodes) and the information on the computing frequency (e.g., once per hour) to obtain a computing requirement sub-feature, which is a data structure containing these three pieces of information.

[0042] Step 144: Extract the decryption features of the encryption gradient data, the decryption features of the encryption business data, and the encryption update features from the encryption state information of the fused encryption dataset.

[0043] In this embodiment, the data processing system extracts decryption features of the encryption gradient data from the encryption status information of the fused encryption dataset. These features include the key type and permission level required for decryption. The system also extracts decryption features of the encrypted business data, including the time and resources required for decryption. Finally, it extracts encryption update features, including the method and triggering conditions for data updates. For example, the decryption features of the encryption gradient data may include the need for a symmetric key and a high permission level; the decryption features of the encrypted business data may include a decryption time of no more than 1 second and required resources of no more than 10MB; and the encryption update features may include periodic updates and a triggering condition where the data change exceeds 10%.

[0044] Step 145: Perform range processing on the decryption features of the encrypted gradient data to generate decryption node range information of the encrypted gradient data; perform path processing on the decryption features of the encrypted business data to generate decryption path length information of the encrypted business data; perform periodic processing on the encrypted update features to generate encrypted update frequency information.

[0045] In this embodiment, the data processing system performs range processing on the decryption characteristics of the encrypted gradient data. For example, based on the permission level required for decryption, it determines the range of nodes capable of decrypting the encrypted gradient data as nodes A, B, and C, generating decryption node range information for the encrypted gradient data. It also performs path processing on the decryption characteristics of the encrypted business data. For example, based on the time and resources required for decryption, it determines the path length required to decrypt the encrypted business data to be 1, generating decryption path length information for the encrypted business data. Finally, it performs periodic processing on the encryption update characteristics. For example, based on the method and triggering conditions for encrypted data updates, it determines the encryption update frequency to be once every two days, generating encryption update frequency information.

[0046] Step 1451: Entity matching is performed on the encryption status information of the fused encryption dataset based on the pre-built decryption feature knowledge graph. The decryption feature knowledge graph includes decryption node entities, decryption permission relationships and feature association rules. Nodes include decryption request initiation nodes, permission control nodes, gradient data ownership nodes and encryption operation nodes. Edges are the decryption permission granting, feature attribute mapping and operation log association relationships between nodes.

[0047] In this embodiment, the data processing system calls a pre-built decryption feature knowledge graph. The node types of this graph cover the roles of the entire decryption process: the decryption request initiating node corresponds to the business node that initiates the decryption application in the cloud office scenario (e.g., node A is the project team document editing node), the permission control node corresponds to the platform permission management module (e.g., node D is the cloud office permission center), the gradient data ownership node corresponds to the federated learning gradient generation node (e.g., nodes A, B, and C are business nodes participating in gradient training), and the encryption operation node corresponds to the technical node that performs the encryption action (e.g., node E is the cloud office encryption service node). The edges of the graph carry the relationship attributes between nodes: the decryption permission granting edge records the rules for permission control node D to grant advanced decryption permissions to business node A; the feature attribute mapping edge records the mapping relationship between the gradient features of gradient data ownership node A and the encryption features of encryption operation node E; and the operation log association edge records the association link between the application log of decryption request initiating node A and the approval log of permission control node D. The data processing system will integrate the encryption status information of the encrypted dataset (such as "the decryption permission of encrypted gradient data X is marked as advanced") into the knowledge graph, and use entity recognition algorithms to locate entities such as "encrypted gradient data X" and "advanced permission" in the information, and perform preliminary matching with the gradient data ownership node and permission control node in the graph.

[0048] Step 1452: Match the fields in the encrypted state information with the graph nodes using the entity linking technology of the decryption feature knowledge graph, and filter out the feature values ​​related to the decryption of the encrypted gradient data.

[0049] In this embodiment, the data processing system employs entity linking technology to precisely match fields in the encrypted status information: For the "decryption permission level is high" field in the encrypted status information, it first locates the "high permission" attribute of the permission control node D in the location graph through string matching, and then confirms the association between this attribute and the gradient data's owner node A (i.e., node A has high decryption permission); for the "encrypted gradient data X is generated by node A" field, it matches the "gradient data X" feature entity of the gradient data's owner node A in the location graph through feature hashing. After matching, feature values ​​directly related to the decryption of the encrypted gradient data are filtered out, including the "high decryption permission identifier" of node A, the "owner node identifier" of gradient data X, and the "approval node identifier" of the permission control node D.

[0050] Step 1453: Perform knowledge graph association reasoning on the selected feature values, traverse the graph nodes and associated edges corresponding to the feature values, and calculate the association strength between the nodes.

[0051] In this embodiment, the data processing system starts with the selected feature values ​​and traverses the associated edges of the knowledge graph for reasoning: starting from the "Advanced Decryption Permission Identifier" feature value of node A, it traverses along the "Decryption Permission Grant" edge to the permission control node D and calculates the weight of the edge (e.g., the weight of the valid option of permission grant is 0.9); starting from the "Origin Node Identifier" feature value of gradient data X, it traverses along the "Feature Attribute Mapping" edge to the encryption operation node E and calculates the weight of the edge (e.g., the accuracy weight of feature mapping is 0.8); starting from the "Approval Node Identifier" feature value of permission control node D, it traverses along the "Operation Log Association" edge to the decryption request initiating node A and calculates the weight of the edge (e.g., the integrity weight of log association is 0.7). The association strength is calculated using a weighted summation method, that is, the association strength between nodes = edge weight × node attribute matching degree. For example, the association strength between nodes A and D = 0.9 (permission granting edge weight) × 0.95 (permission attribute matching degree of node A), and the association strength between nodes A and E = 0.8 (feature mapping edge weight) × 0.9 (feature matching degree).

[0052] Step 1454: Construct a range candidate set of decryption nodes based on association strength, include nodes with association strength that meet the conditions into the core decryption node candidate set, and include nodes with association strength that meet the conditions and the path number requirement into the extended decryption node candidate set.

[0053] In this embodiment, the data processing system presets a core association strength threshold of 0.8, an extended association strength threshold of 0.6, and a path number threshold of 2. First, it filters nodes with an association strength ≥ 0.8. The association strength between nodes A and D is 0.855 ≥ 0.8, the association strength between nodes B and D is 0.82 ≥ 0.8, and the association strength between nodes C and D is 0.81 ≥ 0.8. These nodes are included in the core decryption node candidate set (i.e., candidate set S1 = {A, B, C, D}). Next, it filters nodes with an association strength ≥ 0.6 and a path number ≤ 2. The association strength between encryption operation node E and node A is 0.72 ≥ 0.6, and the path from E to A is E→D→A (the path number is 2), which meets the path requirement. Node E is included in the extended decryption node candidate set (i.e., candidate set S2 = {E}). At the same time, nodes with an association strength < 0.6 are excluded (such as the association strength between nodes F and D being 0.5 < 0.6).

[0054] Step 1455: Determine the range boundary of the candidate set according to the predefined decryption range category of the decryption feature knowledge graph. If the proportion of the number of nodes in the core decryption node candidate set meets the condition and the business scenario correlation between the extended decryption node candidate set and the core node meets the condition, then match the regular decryption range category in the decryption feature knowledge graph. The decryption node range information is the union of the core candidate set and the extended candidate set.

[0055] In this embodiment, the knowledge graph predefined decryption range categories include a regular category (core percentage ≥ 60% and extended correlation ≥ 0.7) and a strict category (core percentage ≥ 60% and extended correlation < 0.7). The data processing system first calculates the percentage of nodes in the core candidate set: the total number of nodes is 6 (A, B, C, D, E, F), and the number of nodes in the core candidate set S1 is 4, with a percentage of 4 / 6 ≈ 66.7% ≥ 60%; then, it calculates the business scenario correlation between the extended candidate set S2 and the core candidate set S1: through business tag matching in the knowledge graph (the correlation between the "encryption service" tag of node E and the "document editing" tag of node A is 0.8), the correlation is 0.8 ≥ 0.7. Therefore, the regular decryption range category is matched, and the decryption node range information is S1∪S2={A, B, C, D, E}.

[0056] Step 1456: If the proportion of nodes in the core decryption node candidate set meets the conditions and the extended candidate set is empty or the business scenario relevance does not meet the conditions, then the strict decryption range category in the decryption feature knowledge graph is matched, and the decryption node range information is the core candidate set.

[0057] In this embodiment, the extended candidate set S2 is empty (i.e., there are no nodes with an association strength ≥ 0.6 and a path count ≤ 2). The data processing system calculates that the node count of the core candidate set still accounts for 66.7% ≥ 60%, but the extended candidate set is empty. Therefore, it matches the strict decryption range category, and the decryption node range information is the core candidate set S1 = {A, B, C, D}. If the extended candidate set S2 = {F} (the association strength of node F is 0.6, but the business scenario association degree with the core node is 0.6 < 0.7), then it also matches the strict category, and the decryption node range information is S1 = {A, B, C, D}.

[0058] Step 146: Combine the decryption node range information, decryption path length information, and encryption update frequency information into encryption requirement sub-features.

[0059] In this embodiment of the application, the data processing system combines the decryption node range information (e.g., nodes A, B, C), the decryption path length information (e.g., 1), and the encryption update frequency information (e.g., once every two days) to obtain the encryption requirement sub-feature, which is a data structure containing these three pieces of information.

[0060] Step 150: Based on the association mapping between the computational requirement sub-features and the encryption requirement sub-features, the current federated learning strategy and the multi-party secure computation protocol are synchronously optimized, so that the gradient calculation process of the optimized federated learning strategy and the encryption processing process of the optimized multi-party secure computation protocol are coordinated.

[0061] In this embodiment, the data processing system first establishes a mapping between computational requirement sub-features and encryption requirement sub-features. For example, it associates the number of computational nodes with the range of decryption nodes, the computational path length with the decryption path length, and the computational frequency with the encryption update frequency. Then, based on these mappings, it synchronously optimizes the current federated learning strategy and the multi-party secure computation protocol. For instance, if the number of computational nodes is 3 and the range of decryption nodes is nodes A, B, and C, it optimizes the gradient aggregation rule of the current federated learning strategy to ensure that the number of nodes in the gradient aggregation is 3; it also optimizes the node interaction order rule of the multi-party secure computation protocol to ensure that the number of nodes in the interaction order is 3. Similarly, based on the mapping between computational path length and decryption path length, it optimizes the gradient propagation path length of the current federated learning strategy and the path length in the privacy-preserving computation logic of the multi-party secure computation protocol; and based on the mapping between computational frequency and encryption update frequency, it optimizes the gradient computation frequency of the current federated learning strategy and the protocol update frequency of the multi-party secure computation protocol. Ultimately, the gradient calculation process of the optimized federated learning strategy is coordinated with the encryption process of the optimized multi-party secure computation protocol. For example, the number of nodes, path length, and frequency in the gradient calculation process are consistent with the number of nodes, path length, and frequency in the encryption process.

[0062] Step 151: Establish an association mapping between the number of computing nodes in the computing requirement sub-feature and the range of decryption nodes in the encryption requirement sub-feature; establish an association mapping between the computing path length in the computing requirement sub-feature and the decryption path length in the encryption requirement sub-feature; and establish an association mapping between the computing frequency in the computing requirement sub-feature and the encryption update frequency in the encryption requirement sub-feature.

[0063] In this embodiment, the data processing system establishes an association mapping between the number of computing nodes (e.g., 3 nodes) and the range of decryption nodes (e.g., nodes A, B, and C), for example, mapping 3 nodes one-to-one with nodes A, B, and C. It also establishes an association mapping between computing path length information (e.g., 2 nodes) and decryption path length information (e.g., 1), for example, associating the computing path length of 2 nodes with the decryption path length of 1, and adjusting the computing path length to 1. Finally, it establishes an association mapping between computing frequency information (e.g., once per hour) and encryption update frequency information (e.g., once every two days), for example, associating the hourly computing frequency with the two-day encryption update frequency, and adjusting the computing frequency to once every two days.

[0064] Step 152: Adjust the gradient aggregation rule of the current federated learning strategy according to the correlation mapping between the number of computing nodes and the range of decryption nodes, so that the number of nodes for gradient aggregation is consistent with the range of decryption nodes; adjust the gradient propagation path of the current federated learning strategy according to the correlation mapping between the computing path length and the decryption path length, so that the gradient propagation path length is consistent with the decryption path length; adjust the gradient update frequency of the current federated learning strategy according to the correlation mapping between the computing frequency information and the encryption update frequency information, so that the gradient update frequency is consistent with the encryption update frequency.

[0065] In this embodiment, the data processing system adjusts the gradient aggregation rules of the current federated learning strategy based on the correlation mapping between the number of computation nodes and the range of decryption nodes. For example, if the number of computation nodes is 3 and the range of decryption nodes is nodes A, B, and C, the number of nodes in the gradient aggregation rules is adjusted to 3, making the number of nodes in the gradient aggregation consistent with the range of decryption nodes. The system also adjusts the gradient propagation path of the current federated learning strategy based on the correlation mapping between the computation path length and the decryption path length. For example, if the computation path length is 2 nodes and the decryption path length is 1, the gradient propagation path length is adjusted to 1, making it consistent with the decryption path length. Finally, the system adjusts the gradient update frequency of the current federated learning strategy based on the correlation mapping between the computation frequency and the encryption update frequency. For example, if the computation frequency is once per hour and the encryption update frequency is once every two days, the gradient update frequency is adjusted to once every two days, making it consistent with the encryption update frequency.

[0066] Step 1521: Extract the initial number of aggregation nodes from the computation node number information of the computation requirement sub-feature; extract the number of decryption nodes from the decryption node range information of the encryption requirement sub-feature; calculate the difference between the initial number of aggregation nodes and the number of decryption nodes; if the difference is positive, reduce the number of nodes in the gradient aggregation rule by an amount equal to the difference; if the difference is negative, increase the number of nodes in the gradient aggregation rule by an amount equal to the absolute value of the difference; perform feature matching processing on the adjusted number of aggregation nodes, matching the adjusted number of nodes with the features of the decryption node range; if the matching degree is higher than the preset matching degree, retain the adjusted aggregation rule; if the matching degree is not higher than the preset matching degree, readjust the number of aggregation nodes.

[0067] In this embodiment, the data processing system extracts the initial number of aggregation nodes from the computation node quantity information of the computation requirement sub-feature, for example, 4 nodes. It extracts the number of decryption nodes from the decryption node range information of the encryption requirement sub-feature, for example, 3 nodes. The difference between the initial number of aggregation nodes and the number of decryption nodes is calculated, i.e., 4-3=1, and the difference is positive. The number of nodes in the gradient aggregation rule is then reduced by an amount equal to the difference, i.e., 1 node is reduced, resulting in an adjusted number of aggregation nodes of 3 nodes. Next, feature matching processing is performed on the adjusted number of aggregation nodes (3 nodes) and matched with the features of the decryption node range (e.g., features of nodes A, B, and C), calculating the matching degree. The preset matching degree is 0.8. If the matching degree is 0.9, which is higher than the preset matching degree, the adjusted aggregation rule is retained; if the matching degree is 0.7, which is not higher than the preset matching degree, the number of aggregation nodes is readjusted, for example, to 2 nodes, and feature matching processing is performed again until the matching degree is higher than the preset matching degree.

[0068] Step 1522: Extract the initial path length from the computation path length information of the computation requirement sub-feature; extract the target path length from the decryption path length information of the encryption requirement sub-feature; calculate the difference between the initial path length and the target path length; if the difference is positive, shorten the gradient propagation path length by equal to the difference; if the difference is negative, extend the gradient propagation path length by equal to the absolute value of the difference; perform state transition processing on the adjusted path length, converting the state of the computation path to the state of the decryption path; if the success rate of the state transition is higher than the preset success rate, retain the adjusted propagation path; if the success rate is not higher than the preset success rate, readjust the path length.

[0069] In this embodiment, the data processing system extracts the initial path length from the computation path length information of the computation requirement sub-feature, for example, 3. It extracts the target path length from the decryption path length information of the encryption requirement sub-feature, for example, 2. The difference between the initial path length and the target path length is calculated, i.e., 3-2=1, and the difference is positive. The gradient propagation path length is then shortened, with the shortened length equal to the difference, i.e., shortened by 1, resulting in an adjusted path length of 2. Next, a state transition process is performed on the adjusted path length, converting the computation path state to the decryption path state, for example, converting the node order in the computation path to the node order in the decryption path. The success rate of the state transition is calculated, for example, 0.9. A preset success rate is 0.8. If the success rate is higher than the preset success rate, the adjusted propagation path is retained; if the success rate is not higher than the preset success rate, the path length is readjusted, for example, the path length is adjusted to 1, and the state transition process is performed again until the success rate is higher than the preset success rate.

[0070] Step 153: Adjust the node interaction order of the multi-party secure computation protocol according to the association mapping between the number of computing nodes and the range of decryption nodes, so that the number of nodes in the node interaction order is consistent with the range of decryption nodes. Adjust the privacy protection computing logic of the multi-party secure computation protocol according to the association mapping between the computing path length information and the decryption path length information, so that the path length in the privacy protection computing logic is consistent with the decryption path length. Adjust the protocol update frequency of the multi-party secure computation protocol according to the association mapping between the computing frequency information and the encryption update frequency information, so that the protocol update frequency is consistent with the encryption update frequency.

[0071] In this embodiment, the data processing system adjusts the node interaction order of the multi-party secure computation protocol based on the correlation mapping between the number of computing nodes and the range of decryption nodes. For example, if the number of computing nodes is 3 and the range of decryption nodes is nodes A, B, and C, the number of nodes in the node interaction order is adjusted to 3, making the number of nodes in the node interaction order consistent with the range of decryption nodes. The privacy-preserving computation logic of the multi-party secure computation protocol is adjusted based on the correlation mapping between the computation path length and the decryption path length. For example, if the computation path length is 2 nodes and the decryption path length is 1, the path length in the privacy-preserving computation logic is adjusted to 1, making the path length in the privacy-preserving computation logic consistent with the decryption path length. The protocol update frequency of the multi-party secure computation protocol is adjusted based on the correlation mapping between the computation frequency and the encryption update frequency. For example, if the computation frequency is once per hour and the encryption update frequency is once every two days, the protocol update frequency is adjusted to once every two days, making the protocol update frequency consistent with the encryption update frequency.

[0072] Step 154: Determine whether the gradient calculation process of the adjusted federated learning strategy and the encryption process of the adjusted multi-party secure computation protocol are coordinated. If they are coordinated, complete the synchronization optimization. If they are not coordinated, repeat the adjustment steps until they are coordinated.

[0073] In this embodiment, the data processing system first extracts features of the gradient calculation process of the adjusted federated learning strategy, including the number of nodes for gradient calculation, the path length for gradient calculation, and the frequency of gradient calculation; it also extracts features of the encryption process of the adjusted multi-party secure computation protocol, including the number of nodes for encryption processing, the path length for encryption processing, and the frequency of encryption processing. Then, it compares whether the above features are consistent. If they are consistent, it determines cooperation and completes synchronization optimization; if they are inconsistent, it determines non-cooperation and repeats the adjustment steps until cooperation is achieved.

[0074] Step 1541: Extract node interaction temporal features, logical execution dependency features, and frequency synchronization features from the gradient calculation process of the adjusted federated learning strategy, and extract the corresponding node interaction temporal features, logical execution dependency features, and frequency synchronization features from the encryption process of the adjusted multi-party secure computation protocol; establish a bidirectional mapping relationship between computation-side features and encryption-side features, generate a feature mapping matrix, where matrix elements represent the mapping strength of corresponding feature pairs; set an upper limit for the number of iteration verification rounds, and initialize the verification weights for each feature dimension; in each round of verification, calculate the weighted mean of each element in the feature mapping matrix, where the weighting coefficient is the verification weight of the corresponding feature dimension; if the weighted mean is lower than the preset value... If the mapping strength threshold is reached, the verification weights of the corresponding feature dimensions are adjusted, increasing the weights of feature dimensions with low mapping strength and decreasing the weights of feature dimensions with high mapping strength. The feature mapping matrix is ​​updated, and the weighted mean is recalculated based on the adjusted weights. The weight adjustment and mean calculation steps are repeated until the change in the weighted mean is less than the preset stability threshold or the number of iterations reaches the upper limit. The weighted mean results of all iterations are combined. If the stable weighted mean is higher than the collaboration judgment threshold, the gradient calculation process and the encryption process are determined to be collaborative. If the stable weighted mean is lower than the collaboration judgment threshold, the process is determined to be uncooperative, and the adjustment steps are returned to re-optimize the federated learning strategy and the multi-party secure computation protocol.

[0075] In this embodiment, the data processing system first extracts node interaction timing features (e.g., the interaction timing of node A→node B→node C), logical execution dependency features (e.g., gradient calculation depends on the result of the previous gradient calculation), and frequency synchronization features (e.g., the gradient calculation frequency is consistent with the encryption update frequency) from the gradient calculation process of the adjusted federated learning strategy; and extracts corresponding node interaction timing features (e.g., the interaction timing of node A→node B→node C), logical execution dependency features (e.g., encryption processing depends on the result of the previous encryption processing), and frequency synchronization features (e.g., the encryption processing frequency is consistent with the gradient calculation frequency) from the encryption processing process of the adjusted multi-party secure computation protocol.

[0076] Then, a bidirectional mapping relationship is established between computation-side features and encryption-side features, generating a feature mapping matrix. Matrix elements represent the mapping strength of corresponding feature pairs. For example, the mapping strength of node interaction timing feature pairs is 0.9, the mapping strength of logic execution dependency feature pairs is 0.8, and the mapping strength of frequency synchronization feature pairs is 0.7. The upper limit of iterative verification rounds is set to 10 times, and the verification weights of each feature dimension are initialized. For example, the weight of node interaction timing features is 0.4, the weight of logic execution dependency features is 0.3, and the weight of frequency synchronization features is 0.3.

[0077] In each round of validation, the weighted average of each element in the feature mapping matrix is ​​calculated. For example, the weighted average in the first round is 0.9×0.4 + 0.8×0.3 + 0.7×0.3. The preset mapping strength threshold is 0.85. If the weighted average is lower than this threshold, the validation weights of the corresponding feature dimensions are adjusted. For example, the weights of logical execution dependency features and frequency synchronization features are increased, while the weight of node interaction timing features is decreased. After adjustment, the weights are 0.3 for node interaction timing features, 0.4 for logical execution dependency features, and 0.3 for frequency synchronization features. The feature mapping matrix is ​​then updated, for example, by recalculating the mapping strength of each feature pair to obtain a new feature mapping matrix.

[0078] The weighted mean is recalculated based on the adjusted weights. For example, the weighted mean in the second round is 0.9×0.3+0.8×0.4+0.7×0.3=0.8. The weight adjustment and mean calculation steps are repeated until the change in the weighted mean is less than a preset stability threshold (e.g., 0.01) or the maximum number of iterations is reached.

[0079] Based on the weighted average results of all iterations, if the stable weighted average is higher than the collaboration decision threshold (e.g., 0.85), the gradient calculation process and the encryption process are determined to be collaborative; if the stable weighted average is lower than the collaboration decision threshold, they are determined to be uncooperative, and the process returns to the adjustment step to re-optimize the federated learning strategy and the multi-party secure computation protocol.

[0080] In an optional embodiment, the method further includes: Step 210: In response to the collaborative result of the gradient calculation process of the optimized federated learning strategy and the encryption processing process of the optimized multi-party secure computation protocol, extract the computation sharding requirement features related to data sharding in the gradient calculation process, and the encryption sharding compatibility features related to data sharding in the encryption processing process; wherein, the computation sharding requirement features include feature granularity information of gradient calculation, cross-sharding dependency information of gradient data, and sharding association strength information of gradient aggregation, and the encryption sharding compatibility features include encryption data block size adaptation information, sharding processing complexity information of encryption algorithm, and sharding dependency path information of decryption operation.

[0081] In this embodiment, the data processing system, responding to the collaborative results, first extracts the computational sharding requirement features related to data sharding during gradient calculation. The feature granularity information for gradient calculation is the size of the data shards during gradient calculation, for example, 100MB; the cross-shard dependency information for gradient data is the dependency relationship between gradient data in different shards, for example, the gradient data of shard 1 depends on the gradient data of shard 2; the shard association strength information for gradient aggregation is the degree of association between different shards, for example, the association strength between shard 1 and shard 2 is 0.8. Next, the encryption sharding compatibility features related to data sharding during encryption processing are extracted. The size adaptation information for encrypted data blocks is the degree of adaptation between the size of the encrypted data blocks and the size of the data shards, for example, 0.9; the sharding processing complexity information for the encryption algorithm is the complexity of the encryption algorithm in processing data shards, for example, low complexity; the sharding dependency path information for decryption operations is the dependency path of the decryption operation between different shards, for example, the path to decrypt shard 1 is shard 1 → shard 2.

[0082] Step 220: Associate and map the computation sharding requirement features with the encryption sharding compatibility features to generate data sharding dynamic adjustment rules. The adjustment rules are used to match the correspondence between computation sharding requirements and encryption sharding compatibility features.

[0083] In this embodiment, the data processing system correlates the gradient calculation feature granularity information in the sharding requirement features with the encrypted data block size adaptation information in the encryption sharding compatibility features; correlates the cross-sharding dependency information of gradient data with the sharding dependency path information of the decryption operation; and correlates the sharding association strength information of gradient aggregation with the sharding processing complexity information of the encryption algorithm. Based on the above correlation mappings, dynamic data sharding adjustment rules are generated. For example, if the gradient calculation feature granularity is 100MB and the encrypted data block size adaptation information is 0.9, then the data sharding size is adjusted to 100MB; if the cross-sharding dependency relationship of gradient data is shard 1 depends on shard 2, and the sharding dependency path of the decryption operation is shard 1 → shard 2, then this dependency relationship is maintained; if the sharding association strength of gradient aggregation is 0.8 and the sharding processing complexity of the encryption algorithm is low, then this association strength is maintained.

[0084] Step 230: According to the data sharding dynamic adjustment rules, adjust the sharding method of the business dataset of the target business scenario to obtain the adjusted business sharded dataset.

[0085] In this embodiment, the data processing system adjusts the sharding method of the business dataset for the target business scenario according to the data sharding dynamic adjustment rules. For example, if the data sharding dynamic adjustment rules require the data shard size to be adjusted to 100MB, then the document editing operation data, meeting arrangement data, task allocation data, etc. in the business dataset are sharded into 100MB shards to obtain the adjusted business sharded dataset, which is a data set containing multiple 100MB shards.

[0086] Step 240: Use the adjusted business shard dataset as input to the optimized multi-party secure computation protocol, combine it with the gradient data of the optimized federated learning strategy for privacy-preserving fusion processing, and update the fused encrypted dataset.

[0087] In this embodiment, the data processing system inputs the adjusted business shard dataset into the optimized multi-party secure computation protocol and performs privacy-preserving fusion processing by combining the gradient data of the optimized federated learning strategy. For example, shard 1 in the adjusted business shard dataset is fused with the corresponding part of the gradient data of the optimized federated learning strategy, and then encrypted using homomorphic encryption technology to obtain fused encrypted data shard 1. Similarly, fused encrypted data shard 2, fused encrypted data shard 3, etc., are obtained. Then, the fused encrypted dataset is updated based on the above fused encrypted data shards, and the updated fused encrypted dataset is a data set containing all fused encrypted data shards.

[0088] Step 250: Based on the updated fused encrypted dataset, synchronously update the sharding logic in the gradient calculation process of the optimized federated learning strategy, so that the sharding logic of gradient calculation is consistent with the sharding logic of encryption processing.

[0089] In this embodiment, the data processing system synchronously updates the sharding logic in the gradient calculation process of the optimized federated learning strategy based on the updated fused encrypted dataset. For example, if the shard size of the updated fused encrypted dataset is 100MB, the sharding logic in the gradient calculation process is adjusted to shard according to the 100MB size, so that the sharding logic of gradient calculation is consistent with the sharding logic of encryption processing.

[0090] In an optional embodiment, the method further includes: Step 310: In response to the coordinated results of the gradient calculation process of the optimized federated learning strategy and the encryption processing process of the optimized multi-party secure computation protocol, collect gradient anomaly features in the gradient calculation process and encryption anomaly features in the encryption processing process; wherein, the gradient anomaly features include information on the deviation of gradient value change trend, information on node interaction delay in gradient transmission, and information on consistency deviation of gradient aggregation results; the encryption anomaly features include information on integrity verification failure of encrypted data, information on permission verification anomaly of decryption operation, and information on temporal misalignment of encryption update.

[0091] In this embodiment, the data processing system, in response to the collaborative results, first collects gradient anomaly features during the gradient calculation process. The gradient value change trend deviation information is the degree of deviation between the gradient value change trend and the expected change trend, for example, 0.2; the gradient transmission node interaction delay information is the delay time for gradient transmission between nodes, for example, 1 second; the gradient aggregation result consistency deviation information is the degree of deviation between the gradient aggregation result and the expected result, for example, 0.1. Next, it collects encryption anomaly features during the encryption processing process. The encryption data integrity verification failure information is the number of times the encryption data integrity verification failed, for example, 1 time; the decryption operation permission verification anomaly information is the number of times the decryption operation permission verification failed, for example, 0 times; the encryption update timing misalignment information is the degree of misalignment between the encryption update timing and the expected timing, for example, 0.1.

[0092] Step 320: Perform correlation analysis between the gradient anomaly features and the encryption anomaly features to locate the root cause of the collaboration mismatch. The root cause of the collaboration mismatch includes the dimensional mismatch between the gradient calculation features and the encryption processing features, the rhythm mismatch between the gradient calculation time sequence and the encryption processing time sequence, and the permission mismatch between the gradient calculation node and the encryption processing node.

[0093] In this embodiment, the data processing system performs correlation analysis on gradient anomaly features and encryption anomaly features. For example, it analyzes the correlation between gradient value trend deviation information and encrypted data integrity verification failure information; it analyzes the correlation between gradient transmission node interaction delay information and decryption operation permission verification anomaly information; and it analyzes the correlation between gradient aggregation result consistency deviation information and encryption update timing misalignment information. Based on the above correlation analysis, the root cause of the collaboration mismatch leading to the anomaly is located. For example, if the gradient value trend deviation information and encrypted data integrity verification failure information have a high correlation, the root cause of the collaboration mismatch is a dimensional mismatch between gradient calculation features and encryption processing features; if the gradient transmission node interaction delay information and decryption operation permission verification anomaly information have a high correlation, the root cause of the collaboration mismatch is a permission mismatch between gradient calculation nodes and encryption processing nodes; if the gradient aggregation result consistency deviation information and encryption update timing misalignment information have a high correlation, the root cause of the collaboration mismatch is a rhythm mismatch between gradient calculation timing and encryption processing timing.

[0094] Step 330: Based on the root cause of the collaborative mismatch, generate a corresponding self-repair strategy, which is used to adjust the gradient calculation features of the optimized federated learning strategy or the encryption processing features of the optimized multi-party secure computation protocol.

[0095] In this embodiment, the data processing system generates a corresponding self-repair strategy based on the root cause of the collaboration mismatch. For example, if the root cause of the collaboration mismatch is a mismatch in the dimensions of the gradient calculation features and the encryption processing features, the generated self-repair strategy is to adjust the dimensions of the gradient calculation features of the optimized federated learning strategy to match the dimensions of the encryption processing features; if the root cause of the collaboration mismatch is a mismatch in the timing of the gradient calculation and the timing of the encryption processing, the generated self-repair strategy is to adjust the timing of the gradient calculation of the optimized federated learning strategy to match the timing of the encryption processing; if the root cause of the collaboration mismatch is a mismatch in the permissions of the gradient calculation node and the encryption processing node, the generated self-repair strategy is to adjust the permissions of the encryption processing node of the optimized multi-party secure computation protocol to match the permissions of the gradient calculation node.

[0096] Step 340: Apply the self-healing strategy to the optimized federated learning strategy and the optimized multi-party secure computation protocol to update the gradient calculation process and the encryption process.

[0097] In this embodiment, the data processing system applies a self-repair strategy to the optimized federated learning strategy and the optimized multi-party secure computation protocol. For example, if the self-repair strategy is to adjust the dimension of the gradient calculation features of the optimized federated learning strategy, then the dimension of the gradient calculation features is adjusted from 3D to 2D, and the gradient calculation process is updated; if the self-repair strategy is to adjust the permissions of the encryption processing nodes of the optimized multi-party secure computation protocol, then the permissions of the encryption processing nodes are adjusted from high-level to medium-level, and the encryption processing process is updated.

[0098] Step 350: Based on the updated gradient calculation process and encryption process, re-execute the collaborative verification of the gradient calculation and encryption process to keep the updated gradient calculation process and encryption process in a collaborative state.

[0099] In this embodiment, the data processing system re-executes the collaborative verification of gradient calculation and encryption processing based on the updated gradient calculation and encryption processes. The verification process is similar to step 1541, extracting features from the updated gradient calculation and encryption processes, establishing a feature mapping matrix, calculating the weighted average, and determining whether they are collaborative. If they are collaborative, the current state is maintained; if they are not collaborative, steps 320 to 350 are repeated until they are collaborative.

[0100] This application's embodiments achieve deep integration of federated learning and multi-party secure computation through a full-link collaborative logic involving gradient update feature-driven multi-party secure computation protocol generation, lightweight encryption fusion, requirement sub-feature extraction, and dual-strategy synchronous optimization. First, an adapted multi-party secure computation protocol is generated based on the gradient update features of the current federated learning strategy, ensuring that the protocol's node interaction order, privacy protection logic, and update frequency accurately match the gradient characteristics of federated learning. Second, lightweight encryption interaction completes the privacy-preserving fusion of gradient data and business datasets, effectively integrating data value while ensuring data privacy. Third, by extracting computation requirement sub-features and encryption requirement sub-features, an association mapping relationship is constructed, providing a quantitative basis for dual-strategy synchronous optimization. Finally, based on this mapping relationship, the federated learning strategy and the multi-party secure computation protocol are synchronously optimized, enabling the gradient computation process of the optimized federated learning strategy and the encryption processing process of the optimized multi-party secure computation protocol to collaborate.

[0101] This design breaks through the bottlenecks of efficiency loss and high privacy leakage risk caused by the independent operation of traditional federated learning and multi-party secure computation. While ensuring data privacy, it improves the efficiency of data fusion and the accuracy of model optimization, thereby achieving secure and efficient data processing in the target business scenario.

[0102] Under another design approach, the data fusion and collaborative optimization method described in this application embodiment can also be applied to the financial and tax management scenario of additional deduction for enterprise R&D expenses. In this scenario, multiple independently accounted subsidiaries within a group each hold their own sensitive financial and tax data, such as R&D project initiation data, personnel hour input, material and equipment procurement records, and patent application information. Due to the involvement of trade secrets and tax compliance requirements, the subsidiaries cannot directly share the original data, but they still need to jointly calculate the group's overall additional deduction amount for R&D expenses. At this time, the target business scenario becomes a collaborative calculation scenario for R&D expenses among multiple legal entities. The data fusion request is initiated by the group's tax management department, and the business dataset includes R&D personnel salary data of subsidiary A, R&D material requisition data of subsidiary B, and external R&D contract data of subsidiary C. The gradient update characteristics of the current federated learning strategy are reflected in the adjustment direction and magnitude of the R&D expense aggregation caliber of each subsidiary.

[0103] In the aforementioned R&D expense management scenario, a multi-party secure computation protocol is used to protect the original financial and tax data of each subsidiary from leakage, while a federated learning strategy is used to continuously optimize the R&D expense allocation model. Specifically, the transmission path in the gradient update features corresponds to the process of R&D expense data flowing from the subsidiary to the group's tax management node. The multi-party secure computation protocol stipulates that subsidiaries interact according to the relevance of R&D projects; for example, subsidiaries related to the same R&D project exchange encrypted expense features first. Lightweight encrypted interaction weights and merges gradient data such as R&D personnel salary weights, material consumption weights, and external R&D expense weights with the original business data of each subsidiary to generate a fused encrypted dataset. The data processing system further extracts computational requirement sub-features (such as the number of subsidiaries participating in expense collection, expense allocation path length, and quarterly collection frequency) and encryption requirement sub-features (such as the scope of tax audit nodes for decrypting R&D expense data, decryption path length, and encryption update cycle), and synchronously optimizes the expense allocation rules in the federated learning strategy and the privacy protection logic in the multi-party secure computation protocol through association mapping. Ultimately, this ensures that the R&D expense calculation process and encryption process of each subsidiary are consistent and coordinated, achieving secure and efficient group R&D expense deduction declaration.

[0104] Understandably, this application scenario also involves anomaly detection and self-repair mechanisms to address collaboration failures caused by changes in fiscal and tax policies or fluctuations in data quality. When adjustments to the R&D expense deduction policy cause deviations in the expected trend of gradient values, or when the integrity verification of encrypted R&D data of one subsidiary fails, the system collects gradient anomaly characteristics (such as deviations in the R&D expense aggregation ratio, data transmission delays between subsidiaries, and inconsistencies in expense aggregation results) and encryption anomaly characteristics (such as encrypted data verification failures, abnormal permission verification, and misaligned encryption update timing). Through correlation analysis, the system identifies the root cause of collaboration mismatches, such as a mismatch between the expense allocation dimension in gradient calculation characteristics and the decryption permission dimension in encryption processing characteristics, or a mismatch between the expense calculation rhythm and the encryption update rhythm. Based on this, the system generates self-repair strategies, such as adjusting the expense aggregation feature dimension in the federated learning strategy to match the decryption node range of the encryption protocol, or renegotiating the encryption update frequency of each subsidiary to align with the quarterly aggregation frequency of R&D expenses, thereby continuously ensuring the compliance, security, and efficiency of data fusion in fiscal and tax management scenarios.

[0105] Please see Figure 2 The figure is a schematic diagram of the basic structure of a data processing system 200 provided in an embodiment of this application. The data processing system 200 includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the data processing methods based on the fusion of multi-party secure computation and federated learning.

[0106] Please see Figure 3 This application provides a functional block diagram of a data processing device, which includes: The multi-source information acquisition module is used to respond to a data fusion request for a target business scenario and acquire the business dataset of the target business scenario and the gradient update features of the current federated learning strategy. The computation protocol generation module is used to generate a multi-party secure computation protocol based on the gradient update features. The multi-party secure computation protocol includes node interaction order rules for protocol execution, privacy-preserving computation logic rules, and protocol update frequency rules. The privacy-preserving fusion module is used to take the gradient data of the current federated learning strategy as the input of the multi-party secure computation protocol, and perform privacy-preserving fusion processing on the gradient data and the business dataset through lightweight encrypted interaction to obtain a fused encrypted dataset. The requirement feature extraction module is used to extract the computation requirement sub-features corresponding to the gradient training features of the current federated learning policy during the policy training phase, as well as the encryption requirement sub-features of the fused encrypted dataset; wherein, the computation requirement sub-features include information on the number of computation nodes for gradient aggregation, information on the length of the computation path for gradient propagation, and information on the frequency of gradient computation, and the encryption requirement sub-features include information on the range of decryption nodes for encrypted gradient data, information on the length of the decryption path for encrypted business data, and information on the encryption update frequency. The synchronous optimization processing module is used to synchronously optimize the current federated learning strategy and the multi-party secure computation protocol based on the association mapping between the computational requirement sub-features and the encryption requirement sub-features, so that the gradient calculation process of the optimized federated learning strategy and the encryption processing process of the optimized multi-party secure computation protocol are coordinated.

[0107] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0108] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of the data processing system reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the data processing system to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0109] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A data processing method based on the fusion of multi-party secure computation and federated learning, characterized in that, include: In response to a data fusion request for a target business scenario, the business dataset of the target business scenario and the gradient update features of the current federated learning strategy are obtained. A multi-party secure computation protocol is generated based on the gradient update features. The multi-party secure computation protocol includes rules for the node interaction order during protocol execution, rules for privacy-preserving computation logic, and rules for protocol update frequency. The gradient data of the current federated learning strategy is used as the input of the multi-party secure computation protocol. The gradient data and the business dataset are fused through a lightweight encrypted interaction to obtain a fused encrypted dataset. Extract the computational requirement sub-features corresponding to the gradient training features of the current federated learning policy during the policy training phase, as well as the encryption requirement sub-features of the fused encrypted dataset; wherein, the computational requirement sub-features include information on the number of computation nodes for gradient aggregation, information on the length of the computation path for gradient propagation, and information on the frequency of gradient computation, and the encryption requirement sub-features include information on the range of decryption nodes for encrypted gradient data, information on the length of the decryption path for encrypted business data, and information on the encryption update frequency. Based on the association mapping between the computational requirement sub-features and the encryption requirement sub-features, the current federated learning strategy and the multi-party secure computation protocol are synchronously optimized, so that the gradient calculation process of the optimized federated learning strategy and the encryption processing process of the optimized multi-party secure computation protocol are coordinated.

2. The method as described in claim 1, characterized in that, The step involves using the gradient data of the current federated learning policy as input to the multi-party secure computation protocol, and performing privacy-preserving fusion processing on the gradient data and the business dataset through lightweight encrypted interaction to obtain a fused encrypted dataset, including: The gradient propagation path in the gradient update features is correlated with the feature values ​​of the gradient data of the current federated learning policy to generate a gradient feature correlation matrix. Perform business feature association processing on the business type information and feature values ​​corresponding to the business dataset of the target business scenario to generate a business feature association matrix; Based on the association results of the gradient feature association matrix and the business feature association matrix, a lightweight encryption method is selected for the encryption interaction: if the feature overlap in the association results is higher than the preset overlap, then a symmetric encryption method is used to encrypt the gradient data and the business dataset; if the feature overlap is lower than the preset overlap, then an asymmetric encryption method is used to encrypt the gradient data and the business dataset. The feature values ​​of the encrypted gradient data are aggregated with the feature values ​​of the encrypted business data to generate aggregated encrypted features. A fused encrypted dataset is then generated based on the aggregated encrypted features. The feature association process and the encryption method selection process are carried out simultaneously, and each adjustment of the association result will affect the encryption method selection logic.

3. The method as described in claim 1, characterized in that, The feature extraction process employs branching logic. The extraction of computational requirement sub-features corresponding to the gradient training features of the current federated learning policy during the policy training phase, and the encryption requirement sub-features of the fused encrypted dataset, includes: Extract gradient aggregation features, gradient propagation features, and gradient frequency features from the gradient training features of the current federated learning strategy; The gradient aggregation features are statistically processed to generate information on the number of computation nodes for gradient aggregation; the gradient propagation features are sequence-processed to generate information on the length of the computation path for gradient propagation; and the gradient frequency features are periodically-processed to generate information on the frequency of gradient computation. The information on the number of computing nodes, the length of computing paths, and the frequency of computing are combined into a sub-feature of computing requirements. Extract the decryption features of the encryption gradient data, the decryption features of the encryption business data, and the encryption update features from the encryption state information of the fused encryption dataset; The decryption features of the encrypted gradient data are processed to generate decryption node range information of the encrypted gradient data; the decryption features of the encrypted business data are processed to generate decryption path length information of the encrypted business data; and the encryption update features are processed periodically to generate encryption update frequency information. The decryption node range information, decryption path length information, and encryption update frequency information are combined into encryption requirement sub-features.

4. The method as described in claim 3, characterized in that, The step of statistically processing the gradient aggregation features to generate information on the number of computation nodes for gradient aggregation includes: Select feature values ​​related to gradient aggregation from the gradient training features of the current federated learning strategy; Statistical analysis is performed on the selected feature values ​​to calculate their distribution, and the number of computation nodes for gradient aggregation is determined based on the distribution of the feature values. If the feature value distribution in the statistical analysis results belongs to the first distribution label, then the number of nodes is calculated as the number of nodes in a concentrated distribution; if the feature value distribution belongs to the second distribution label, then the number of nodes is calculated as the number of nodes in a dispersed distribution. The number of generated computing nodes is dynamically adjusted, and the size of the statistical window for the next statistical iteration is adjusted based on the previous statistical results. Repeat the statistical analysis and dynamic adjustment steps until the number of calculation nodes meets the preset requirements; the statistical processing and dynamic adjustment processes are carried out simultaneously, and each adjustment of the statistical results affects the size of the statistical window for the next time.

5. The method as described in claim 3, characterized in that, The step of performing range processing on the decryption features of the encrypted gradient data to generate decryption node range information for the encrypted gradient data includes: Entity matching is performed on the encryption status information of the fused encryption dataset based on a pre-built decryption feature knowledge graph. The decryption feature knowledge graph includes decryption node entity decryption permission relationship and feature association rules. Nodes include decryption request initiation node, permission control node, gradient data ownership node, and encryption operation node. Edges are the decryption permission granting feature attribute mapping operation log association relationship between nodes. The entity linking technology of the decryption feature knowledge graph is used to match the fields in the encrypted state information with the graph nodes to filter out feature values ​​related to the decryption of the encrypted gradient data. The selected feature values ​​are used for knowledge graph association reasoning. The graph nodes and associated edges corresponding to the feature values ​​are traversed to calculate the association strength between nodes. Based on the association strength, a range candidate set of decryption nodes is constructed. Nodes with association strength that meet the conditions are included in the core decryption node candidate set, and nodes with association strength that meet the conditions and satisfy the path number requirements are included in the extended decryption node candidate set. Based on the predefined decryption range categories of the decryption feature knowledge graph, the candidate set is used to determine the range boundaries. If the proportion of nodes in the core decryption node candidate set meets the conditions and the business scenario relevance between the extended decryption node candidate set and the core node meets the conditions, then the regular decryption range category in the decryption feature knowledge graph is matched, and the decryption node range information is the union of the core candidate set and the extended candidate set. If the proportion of nodes in the core decryption node candidate set meets the conditions and the extended candidate set is empty or the business scenario relevance does not meet the conditions, then the strict decryption range category in the decryption feature knowledge graph is matched, and the decryption node range information is the core candidate set.

6. The method as described in claim 1, characterized in that, The synchronous optimization of the current federated learning strategy and the multi-party secure computation protocol based on the association mapping between the computational requirement sub-features and the encryption requirement sub-features, so that the gradient calculation process of the optimized federated learning strategy and the encryption processing process of the optimized multi-party secure computation protocol are coordinated, includes: Establish an association mapping between the number of computing nodes in the computing requirement sub-feature and the range of decryption nodes in the encryption requirement sub-feature; establish an association mapping between the computing path length in the computing requirement sub-feature and the decryption path length in the encryption requirement sub-feature; and establish an association mapping between the computing frequency in the computing requirement sub-feature and the encryption update frequency in the encryption requirement sub-feature. The gradient aggregation rules of the current federated learning strategy are adjusted according to the correlation mapping between the number of computing nodes and the range of decryption nodes, so that the number of nodes for gradient aggregation is consistent with the range of decryption nodes. The gradient propagation path of the current federated learning strategy is adjusted according to the correlation mapping between the computing path length and the decryption path length, so that the gradient propagation path length is consistent with the decryption path length. The gradient update frequency of the current federated learning strategy is adjusted according to the correlation mapping between the computing frequency information and the encryption update frequency information, so that the gradient update frequency is consistent with the encryption update frequency. The node interaction order of the multi-party secure computation protocol is adjusted according to the association mapping between the number of computing nodes and the range of decryption nodes, so that the number of nodes in the node interaction order is consistent with the range of decryption nodes. The privacy protection computing logic of the multi-party secure computation protocol is adjusted according to the association mapping between the computing path length information and the decryption path length information, so that the path length in the privacy protection computing logic is consistent with the decryption path length. The protocol update frequency of the multi-party secure computation protocol is adjusted according to the association mapping between the computing frequency information and the encryption update frequency information, so that the protocol update frequency is consistent with the encryption update frequency. Determine whether the gradient calculation process of the adjusted federated learning strategy and the encryption process of the adjusted multi-party secure computation protocol are coordinated. If they are coordinated, complete the synchronization optimization; if they are not coordinated, repeat the adjustment steps until they are coordinated.

7. The method as described in claim 6, characterized in that, The step of adjusting the gradient aggregation rule of the current federated learning strategy based on the correlation mapping between the number of computing nodes and the range of decryption nodes, so that the number of nodes for gradient aggregation is consistent with the range of decryption nodes, includes: Extract the initial number of aggregation nodes from the computation node number information of the computation requirement sub-feature; extract the number of decryption nodes from the decryption node range information of the encryption requirement sub-feature; Calculate the difference between the initial number of aggregated nodes and the number of decrypted nodes; if the difference is positive, reduce the number of nodes in the gradient aggregation rule by an amount equal to the difference; if the difference is negative, increase the number of nodes in the gradient aggregation rule by an amount equal to the absolute value of the difference. The adjusted number of aggregated nodes is subjected to feature matching processing, and the adjusted number of nodes is matched with the features of the decrypted node range; If the matching degree is higher than the preset matching degree, the adjusted aggregation rule is retained; if the matching degree is not higher than the preset matching degree, the number of aggregation nodes is readjusted. The step of adjusting the gradient propagation path of the current federated learning policy based on the correlation mapping between the computation path length information and the decryption path length information, so that the gradient propagation path length is consistent with the decryption path length, includes: Extract the initial path length from the computation path length information of the computation requirement sub-feature; extract the target path length from the decryption path length information of the encryption requirement sub-feature; Calculate the difference between the initial path length and the target path length; if the difference is positive, shorten the gradient propagation path by the length equal to the difference; if the difference is negative, extend the gradient propagation path by the length equal to the absolute value of the difference. The adjusted path length undergoes a state transition process, converting the calculated path state to the decrypted path state. If the success rate of the state transition is higher than the preset success rate, the adjusted transmission path is retained; if the success rate is not higher than the preset success rate, the path length is readjusted.

8. The method as described in claim 6, characterized in that, The determination of whether the gradient calculation process of the adjusted federated learning strategy and the encryption process of the adjusted multi-party secure computation protocol are coordinated is as follows: If they are coordinated, the synchronization optimization is completed; if they are not coordinated, the adjustment steps are repeated until they are coordinated. This includes: Extract node interaction timing features, logical execution dependency features, and frequency synchronization features from the gradient calculation process of the adjusted federated learning strategy, and extract the corresponding node interaction timing features, logical execution dependency features, and frequency synchronization features from the encryption process of the adjusted multi-party secure computation protocol. Establish a bidirectional mapping relationship between computation-side features and encryption-side features, and generate a feature mapping matrix, where the matrix elements represent the mapping strength of corresponding feature pairs; Set an upper limit for the number of iteration verification rounds and initialize the verification weights for each feature dimension; in each round of verification, calculate the weighted mean of each element in the feature mapping matrix, and the weighting coefficient is the verification weight of the corresponding feature dimension; If the weighted mean is lower than the preset mapping strength threshold, the verification weight of the corresponding feature dimension will be adjusted, increasing the weight of feature dimensions with low mapping strength and decreasing the weight of feature dimensions with high mapping strength. Update the feature mapping matrix, recalculate the weighted mean based on the adjusted weights, and repeat the weight adjustment and mean calculation steps until the change in the weighted mean is less than the preset stability threshold or the number of iterations reaches the upper limit. By combining the weighted average results of all iterations, if the stabilized weighted average is higher than the collaboration decision threshold, the gradient calculation process and the encryption process are determined to be collaborative; if the stabilized weighted average is lower than the collaboration decision threshold, they are determined to be uncooperative, and the process returns to the adjustment step to re-optimize the federated learning strategy and the multi-party secure computation protocol.

9. A data processing system, characterized in that, include: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the data processing method based on the fusion of multi-party secure computation and federated learning as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the data processing method based on the fusion of multi-party secure computation and federated learning as described in any one of claims 1-8.

Images