Cross-institutional attestation data privacy processing method based on federated learning

Abstract

The application relates to the technical field of federal learning and audit data processing, and discloses a cross-institutional confirmation data privacy processing method based on federal learning, which comprises the following steps: extracting a double-channel edge feature vector from local confirmation response data of each institution and performing anonymization processing; generating a cross-institutional confirmation correlation topology index through a privacy set intersection operation; running a double-channel graph neural network locally in each institution and outputting a node embedding vector through a cross-channel attention gate; performing federal aggregation after applying adaptive differential privacy noise and sparse projection gradient processing; weighting and splicing a global fusion embedding matrix according to a neighborhood size normalization coefficient after privacy protection of the node embedding; calculating a field bias degree measurement value and a delay synergy degree measurement value and fusing them into a joint abnormal edge weight; performing hop-by-hop cumulative propagation to identify an abnormal path; and generating a cross-institutional confirmation data privacy processing result report.

Description

Technical Field

[0001] This invention relates to the field of federated learning and audit data processing technology, specifically to a method for handling cross-agency confirmation data privacy based on federated learning. Background Technology

[0002] In cross-agency confirmation engagements, the auditee typically involves a parent company and multiple subsidiaries, with frequent accounts receivable and payable transactions between these entities. Auditors need to conduct cross-agency comparisons and consistency analyses of the confirmation response data held by each entity to identify discrepancies in financial data and anomalies in response timing.

[0003] Existing privacy-handling methods based on federated learning have two shortcomings when dealing with the above scenarios. Firstly, the related-party transaction topology in cross-institutional confirmation data constitutes core business secrets of the corporate group, and the auditee's identifier and its relationships are sensitive information for audit operations. However, existing federated learning frameworks require sharing plaintext identifiers or complete transaction relationship diagrams when establishing cross-institutional relationships, leading to the risk of leakage of related-party transaction topology and audit task attribution information. Secondly, related parties involved in coordinated fraud simultaneously generate anomalous signals in financial data fields and confirmation response timing. However, these two types of signals are scattered across multi-hop transaction chains and multi-institutional response sequences, respectively. Each dimension, when processed individually, cannot be effectively identified due to the weakness of the signals. Existing federated learning methods lack the ability to perform topological coupling and joint processing of multi-dimensional heterogeneous confirmation data while protecting privacy, making it difficult to effectively detect coordinated fraud anomalies. Summary of the Invention

[0004] This invention provides a privacy processing method for cross-institutional confirmation data based on federated learning, which solves the technical problems in related technologies such as the risk of leakage of original identification information during joint analysis of cross-institutional confirmation data, the difficulty in coordinating the use of local data differences and time delay features of each institution, and the insufficient accuracy of cross-institutional abnormal association path identification.

[0005] This invention discloses a cross-institutional confirmation data privacy processing method based on federated learning, including: Each participating institution extracts edge feature vectors and time delay features from local confirmation response data and concatenates them into dual-channel edge feature vectors. It then anonymizes the counterparty identifier and the auditee identifier to generate anonymous node codes and anonymous association identifiers, and assembles local transaction relationship adjacency fragments. The federal coordination node performs a privacy set intersection operation on the anonymous association identifiers uploaded by each agency to generate a cross-agency confirmation association topology index; Each institution inputs the local transaction relationship adjacency fragment and the dual-channel edge feature vector into the local dual-channel graph neural network, performs graph convolution operations through the data difference message passing channel and the time delay message passing channel respectively, performs feature cross-enhancement through cross-channel attention gating, and outputs the dual-channel local node embedding vector; An adaptive differential privacy noise injection based on node degree is applied to the dual-channel local node embedding vector. After sparse projection processing of the training gradient, calibrated Laplacian noise is added. The gradient is safely aggregated through federated aggregation nodes and iteratively trained until convergence. Based on the privacy protection of each institution through anonymous node encoding, the dual-channel local node embedding vectors are aligned across institutions, and then weighted and concatenated according to the neighborhood scale normalization coefficient to generate a global dual-channel fused node embedding matrix. For each node pair, calculate the field deviation metric and the delay collaboration metric, and map them to the joint anomaly edge weights through a nonlinear fusion function to generate a joint anomaly weighted association graph. Perform hop-by-hop cumulative propagation on the joint anomaly weighted association graph to identify paths where the cumulative joint anomaly score exceeds an adaptive threshold, and generate privacy-preserving results for cross-agency confirmation data.

[0006] Furthermore, the edge feature vector includes transaction amount range encoding, transaction frequency, and transaction type encoding, wherein the transaction amount range encoding and transaction type encoding are converted into numerical representation using one-hot encoding, and the transaction frequency is processed using Z-score normalization; the time delay feature is generated by encoding the response time interval of the confirmation after Z-score normalization; the anonymous node encoding is generated by mapping the counterparty identifier using a one-way hash function, and the anonymous association identifier is generated by encoding the audited entity identifier using locality-sensitive hashing.

[0007] Furthermore, the calculation process for the cross-channel attention gating includes: assuming the data difference channel is in the [missing information]... Layer output nodes The embedding is The timing delay channel is in the first Layer output nodes The embedding is ,in This refers to the layer index of the graph convolutional layer. For node indexing; interaction attention weights are: ; in This is a learnable attention weight vector. For biased scalars, It is the sigmoid activation function. This is a vector concatenation operation; feature cross enhancement is as follows: ; ; in and These are the enhanced data difference channel node embedding and the timing delay channel node embedding, respectively.

[0008] Furthermore, the cross-channel attention gating adopts a multi-head attention approach, which... and Each part is equally divided along its feature dimension into Subvectors, of which Given the number of attention heads, calculate independent attention weights for each head and perform feature cross-enhancement, then concatenate the outputs of each head to restore the original structure.

[0009] Furthermore, the adaptive differential privacy noise injection based on node degree includes: setting the node... The degree in the adjacency segment of the local transaction relation is , This represents the maximum degree of all nodes in the local graph. Based on the noise scale parameter, noise scaling factor for: ; in Indexing nodes; indexing nodes Dual-channel local node embedding vector Embedding with added noise ,in From The noise vectors are independently sampled from the distribution; the sparse projection processing includes preserving the gradients whose absolute values ​​are listed first. The proportion is set to one component and the remaining components are set to zero, where For the sparsity parameter, calibrated Laplacian noise is added to the retained non-zero gradient components.

[0010] Furthermore, the weighted concatenation based on neighborhood-scale normalization coefficients includes: setting anonymous nodes. In the The number of neighbor nodes in each institution is , For containing anonymous nodes The total number of institutions, and the neighborhood size normalization coefficient are: ; in For anonymous node indexing, For institutional indexing, Traversing the index for the organization in the summation; anonymous nodes The global dual-channel fusion node embedding vector is: ; in Anonymous node In the The privacy-preserved dual-channel local node embedding vector in each institution; for anonymous nodes that appear only in a single institution, their global dual-channel fused node embedding vector is the corresponding privacy-preserved dual-channel local node embedding vector in that institution.

[0011] Furthermore, the field deviation metric and the delay co-operation metric are the L2 norm distances between the data difference channel sub-vectors and the temporal delay channel sub-vectors, respectively; the nonlinear fusion function is a multilayer perceptron, with the input being a two-dimensional vector composed of the field deviation metric and the delay co-operation metric, and the output layer connected to a sigmoid activation function, outputting joint anomaly edge weights; when the sets of mechanisms appearing in the two nodes of a node pair are not completely identical, zero vectors are filled into the corresponding channel embedding positions of the missing mechanisms.

[0012] Furthermore, the hop-by-hop cumulative propagation includes: assuming a path sequentially passes through nodes... The cumulative joint anomaly score for this path is: ; in This represents the number of nodes contained in the path. This represents the traversal index of the edges in the path. The joint anomaly edge weights are defined as the weights of adjacent nodes; the adaptive threshold is determined based on the cumulative distribution of the joint anomaly edge weights of historical normal association groups. ; in Let be the mean of this distribution. Standard deviation, This is a threshold sensitivity parameter; for cumulative joint anomaly scores exceeding... When multiple paths have sub-path inclusion relationships, the longest path is retained and its sub-paths are removed.

[0013] Furthermore, the hop-by-hop cumulative propagation employs a pruning strategy: during the hop-by-hop search process, when the accumulated anomaly score from the starting node to the current node plus the estimated value of the maximum joint anomaly edge weight among the remaining reachable nodes multiplied by the maximum remaining hop count is less than an adaptive threshold... If this happens, the search for that path will be terminated.

[0014] This invention discloses a cross-institutional confirmation data privacy processing system based on federated learning, comprising: The local feature extraction and anonymization module is used to extract edge feature vectors and time delay features from local confirmation response data and concatenate them into dual-channel edge feature vectors. It anonymizes the counterparty identifier and the auditee identifier to generate anonymous node codes and anonymous association identifiers, and assembles local transaction relationship adjacency fragments. The cross-institutional association topology generation module is used to receive anonymous association identifiers uploaded by various institutions, perform privacy set intersection operation to identify the set of anonymous association identifiers shared by the institutions, generate a cross-institutional confirmation association topology index and distribute it back to the relevant institutions. The local dual-channel graph neural network processing module is used to aggregate association groups based on the cross-institutional confirmation association topology index, input the local transaction relationship adjacency fragments and dual-channel edge feature vectors into the local dual-channel graph neural network, process the corresponding edge feature sub-vectors through the data difference message passing channel and the time delay message passing channel respectively, perform feature cross-enhancement through cross-channel attention gating, and output dual-channel local node embedding vectors. The differential privacy protection and federated aggregation module is used to apply adaptive differential privacy noise injection based on node degree to the dual-channel local node embedding vector, perform sparse projection processing on the training gradient and add calibrated Laplacian noise, perform safe aggregation of gradients on federated aggregation nodes and distribute global aggregated gradients, and output the trained local dual-channel graph neural network and the privacy-protected dual-channel local node embedding vector set after multiple rounds of iteration. The cross-institutional node alignment and fusion module is used to perform cross-institutional node alignment on the privacy-protected dual-channel local node embedding vector set uploaded by each institution based on anonymous node encoding, and generate a global dual-channel fused node embedding matrix by weighted concatenation according to the neighborhood scale normalization coefficient. The joint anomaly weighted association graph generation module is used to calculate the field deviation metric and the delay collaboration metric for each pair of nodes with transaction relationships in the global dual-channel fusion node embedding matrix, and map them to the joint anomaly edge weights through a nonlinear fusion function to generate a joint anomaly weighted association graph. The abnormal path identification module is used to perform hop-by-hop cumulative propagation calculation on the joint anomaly weighted association graph, accumulate joint anomaly edge weights along connected paths, identify the longest connected path whose cumulative joint anomaly score exceeds an adaptive threshold, and generate a set of dual-channel joint differential propagation paths and the cumulative joint anomaly score of each path. The results report generation module is used to map anonymous nodes on each path in the dual-channel joint differential propagation path set back to the corresponding institutional confirmation records, annotate the main deviation fields and delay mode features, sort them in descending order by cumulative joint anomaly score, and generate a cross-institutional confirmation data privacy processing results report.

[0015] The present invention has the following beneficial effects: First, the counterparty identifier and the auditee identifier are anonymized by one-way hash function and locality-sensitive hash, respectively. Then, a cross-institutional association topology index is established on the federated coordination node by using privacy set intersection operation. Each institution does not need to share plaintext identifiers or complete transaction relationship graphs, thereby avoiding the exposure of related party transaction topology and audit task attribution information during the federated learning process.

[0016] Second, each institution completes the message passing computation of the dual-channel graph neural network locally. It applies adaptive differential privacy noise based on node degree to the embedding vector of the dual-channel local node and performs sparse projection differential privacy processing on the gradient before uploading it to the federated aggregation node. This ensures that the transaction relationship structure and the statistical features of the associated groups do not leave the local institution, thereby protecting the original data and graph structure information of each institution during the federated aggregation and gradient passing process.

[0017] Third, cross-channel attention gating enables feature cross-enhancement of data difference signals and time-delay signals on a shared topology. The learnable nonlinear fusion function enables superlinear enhancement when the two types of signals co-occur. In the scenario of collaborative fraud, weak abnormal signals scattered in multi-hop transaction chains and multi-institution response sequences are effectively amplified after fusion and hop-by-hop cumulative propagation, which helps to solve the problem that single-dimensional processing makes it difficult to identify collaborative anomalies due to weak signals. Attached Figure Description

[0018] Figure 1 This is a flowchart of a cross-institutional confirmation data privacy processing method based on federated learning provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the timing delay and transaction frequency distribution of local confirmation response data of Institution B provided in an embodiment of the present invention; Figure 3 This is a diagram illustrating the number of organizations where each anonymous association identifier appears, based on the cross-organizational privacy set intersection results provided in this embodiment of the invention. Figure 4 This is a schematic diagram comparing the degree of each anonymous node of organization B with the noise scaling factor provided in the embodiment of the present invention; Figure 5 This is a schematic diagram of the normalized coefficient of the cross-institutional embedded fusion neighborhood size of node H3A9F2 provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the comparison of joint abnormal edge weights for typical node pairs provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the hop-by-hop cumulative joint anomaly score of the main abnormal path provided in the embodiments of the present invention; Figure 8This is a schematic diagram of the scatter distribution of the influence of the dual-channel abnormal signal on the joint abnormal edge weight provided in the embodiment of the present invention; Figure 9 This is a schematic diagram illustrating the trend of the output data scale change in each processing step provided in the embodiments of the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a cross-institutional confirmation data privacy processing method based on federated learning, see [link to relevant documentation]. Figure 1 This includes the following steps: Step 1: Feature extraction and anonymization of local confirmation response data.

[0021] Obtain the confirmation response data stored locally by each participating institution. Extract edge feature vectors from each transaction record. These edge feature vectors include transaction amount range encoding, transaction frequency, and transaction type encoding. The transaction amount range encoding and transaction type encoding are categorical data and are converted to numerical representation using one-hot encoding. The transaction frequency is numerical data and is processed using Z-score standardization to eliminate dimensional differences. Simultaneously, extract the corresponding confirmation response time interval. The response time interval is a time value and is encoded as a time-series delay feature after Z-score standardization. Concatenate the edge feature vectors with the time-series delay features to generate a dual-channel edge feature vector.

[0022] The counterparty identifier is mapped to an anonymous node code using a one-way hash function. The auditee identifier to which the confirmation belongs is encoded using locality-sensitive hashing to generate an anonymous association identifier. The anonymous node code, the anonymous association identifier, and the dual-channel edge feature vector are assembled into a local transaction relationship adjacency fragment.

[0023] It should be noted that the dual-channel edge feature vector is a composite vector formed by concatenating the edge feature vector and the time-series delay feature along the feature dimension. The edge feature vector corresponds to the data difference channel, used to characterize the statistical features of transaction amount, frequency, and type; the time-series delay feature corresponds to the time-series delay channel, used to characterize the time interval distribution of confirmation responses. The feature sub-vectors of the two channels are then fed into their respective message passing channels in the subsequent local dual-channel graph neural network processing.

[0024] It should be noted that anonymous node encoding is a fixed-length code obtained by mapping the counterparty identifier using a one-way hash function. The irreversibility of one-way hash functions makes it impossible to deduce the original counterparty identifier from the anonymous node encoding. Anonymous association identifiers are codes generated by applying locality-sensitive hashing to the auditee's identifier. The characteristics of locality-sensitive hashing allow identical or highly similar auditee identifiers to be mapped to the same or similar codes across different institutions, thereby supporting subsequent cross-institutional association matching while avoiding direct exposure of the auditee's identifier.

[0025] Step 2: Generate a cross-institutional association topology index based on the intersection of privacy sets.

[0026] Each institution uploads its anonymous association identifiers to the federal coordination node. The federal coordination node performs a privacy set intersection operation on the anonymous association identifiers from different institutions to identify the set of anonymous association identifiers shared across institutions. The matching results are then distributed back to the relevant institutions in the form of an association index, generating a cross-institutional confirmation association topology index.

[0027] It should be noted that privacy set intersection is a secure multi-party computation protocol where each participating party only knows the intersection elements of their sets and not any information about the non-intersection parts of the other party's sets. In this embodiment, the federated coordination node uses privacy set intersection to match the anonymous association identifiers uploaded by each institution to determine which auditees have confirmation records with multiple institutions simultaneously, while each institution cannot know about auditee information held by other institutions that is unrelated to itself.

[0028] Step 3: Node embedding generation based on local dual-channel graph neural network.

[0029] Each institution aggregates local confirmation response features belonging to the same anonymous association identifier into association groups based on the cross-institutional confirmation association topology index. Local transaction relationship adjacency fragments and dual-channel edge feature vectors are input into a local dual-channel graph neural network, which outputs dual-channel local node embedding vectors.

[0030] Step 3 may specifically include: Step 301: Take the local transaction relationship adjacency fragment as the graph structure input, use anonymous node encoding as the graph node, and use the dual-channel edge feature vector as the edge attribute to initialize the initial embedding representation of each node in the graph.

[0031] Step 302: In the data difference message passing channel of the local dual-channel graph neural network, graph convolution operation is performed on the edge feature vector sub-vectors in the dual-channel edge feature vector based on the adjacency structure, and the data difference feature information of neighboring nodes is aggregated along the transaction relationship edge to generate the data difference channel node embedding.

[0032] Step 303: In the time-delayed message passing channel of the local dual-channel graph neural network, graph convolution operation is performed on the time-delayed feature sub-vectors in the dual-channel edge feature vectors based on the same adjacency structure, and the time-delayed feature information of neighboring nodes is aggregated along the transaction relationship edge to generate the time-delayed channel node embedding.

[0033] Step 304: After each layer of graph convolution, feature cross-enhancement is performed on the data difference channel node embedding and the temporal delay channel node embedding through cross-channel attention gating, and the dual-channel local node embedding vector is output.

[0034] It should be noted that the local dual-channel graph neural network consists of a data difference message passing channel, a temporal delay message passing channel, and a cross-channel attention gating unit. The data difference message passing channel and the temporal delay message passing channel are two parallel graph convolution processing branches. The data difference message passing channel receives the edge feature vector sub-vectors from the dual-channel edge feature vectors as edge attribute inputs, while the temporal delay message passing channel receives the temporal delay feature vectors from the dual-channel edge feature vectors as edge attribute inputs. The two channels share the adjacency structure defined by the local transaction relationship adjacency segment, meaning the message passing paths are the same, but each processes the edge features of its corresponding channel. After each layer of graph convolution is completed, the cross-channel attention gating unit simultaneously receives the node embeddings from both channels, calculates the interaction attention weights between the data difference channel node embeddings and the temporal delay channel node embeddings, and performs weighted fusion enhancement on the node embeddings of the two channels based on these interaction attention weights. The enhanced node embeddings are then fed back to their respective channels as inputs for the next layer of graph convolution. After alternating processing through multiple layers of graph convolution and cross-channel attention gating, the final output is a dual-channel local node embedding vector containing feature information from both channels.

[0035] The local dual-channel graph neural network takes the adjacency segments of local transaction relationships of each institution and the dual-channel edge feature vectors as input, and the dual-channel local node embedding vectors as output. During training, labeled historical confirmation audit data is used, and the labels of known abnormal association groups are used as supervision signals. A linear classification layer is added on top of the dual-channel local node embedding vectors, the loss is calculated using the binary cross-entropy loss function, and the parameters are updated using the Adam optimization algorithm. The training is completed collaboratively by each institution through the federated aggregation process described in step 4.

[0036] It should be noted that the calculation process for cross-channel attention gating is as follows: Let the data difference channel be the... Layer output nodes The embedding is The timing delay channel is in the first Layer output nodes The embedding is Both dimensions are ,in This refers to the layer index of the graph convolutional layer. For node indexing, Embed dimensions for nodes in each channel. The formula for calculating interaction attention weights is: ; in, This is a learnable attention weight vector. Indicates transpose. For biased scalars, It is the sigmoid activation function. This represents a vector concatenation operation. For nodes In the Layer interaction attention weights. The formula for calculating feature cross-enhancement is: ; ; in, and These are the enhanced data difference channel node embedding and timing delay channel node embedding, respectively. When both the data difference signal and the timing delay signal exhibit anomalies at a certain node... As the values ​​approach higher, the anomalous features of the two channels reinforce each other.

[0037] In this embodiment of the application, to improve the expressive power of cross-channel feature cross-enhancement, cross-channel attention gating can adopt a multi-head attention approach. and Each part is equally divided along its feature dimension into Subvectors, of which The number of attention heads, and the dimension of each subvector is [value missing]. After calculating independent attention weights for each head and performing feature cross-enhancement, the outputs of each head are concatenated to restore the original result. Dimensional embedding.

[0038] Step 4: Federated aggregation training based on differential privacy protection.

[0039] An adaptive differential privacy noise injection based on node degree is applied to the dual-channel local node embedding vectors. Sparse projection processing is performed on the training gradients of the local dual-channel graph neural network, projecting the gradients to a low-dimensional sparse subspace and adding calibrated Laplacian noise. The privacy-preserving set of local node embedding vectors and the processed gradients are uploaded to the federated aggregation node. The federated aggregation node performs secure aggregation on the gradients and distributes the globally aggregated gradients to each institution. Each institution updates its local dual-channel graph neural network parameters based on the globally aggregated gradients. After multiple rounds of federated iterations and convergence, the trained local dual-channel graph neural networks of each institution and their corresponding privacy-preserving sets of local node embedding vectors are output.

[0040] Step 4 may specifically include: Step 401: For each node embedding in the dual-channel local node embedding vector, determine the noise scaling factor based on the degree of the corresponding node in the local transaction relationship adjacency segment. Nodes with higher degrees aggregate more neighbor information during message passing, their embeddings contain more structurally sensitive information, and therefore have larger noise scaling factors. Based on the noise scaling factor, sample noise vectors from the Laplace distribution and superimpose them onto the corresponding node's dual-channel local node embedding vector to generate a privacy-preserving dual-channel local node embedding vector.

[0041] Step 402: Perform sparse projection processing on the gradients generated by the local dual-channel graph neural network during the current training round, retaining the gradients with the absolute values ​​listed first. The proportional components are set to zero, while the remaining components are set to zero. is the sparsity parameter. Calibrated Laplacian noise is added to the retained non-zero gradient components to generate the processed gradient.

[0042] In step 403, each institution uploads its privacy-preserving set of local dual-channel node embedding vectors and processed gradients to the federated aggregation node. The federated aggregation node performs secure aggregation operations on the processed gradients uploaded by each institution, calculates the global aggregated gradient, and distributes it to each institution. Each institution updates the parameters of its local dual-channel graph neural network based on the global aggregated gradient.

[0043] Step 404: Repeat steps 301 to 403 until the training loss of the local dual-channel graph neural network converges. After convergence, output the trained local dual-channel graph neural networks of each institution and the set of dual-channel local node embedding vectors after privacy protection for the current round.

[0044] It should be noted that the calculation process for adaptive differential privacy noise injection based on node degree is as follows: Let the node... The degree in the adjacency segment of the local transaction relation is , This represents the maximum degree of all nodes in the local graph. Based on the noise scale parameter, noise scaling factor The calculation formula is: ; in, For node indexing, For nodes The degree, This represents the maximum degree of all nodes in the local graph. This is the basic noise scale parameter. For nodes... Dual-channel local node embedding vector Embedding with added noise ,in From The noise vector sampled independently in the distribution has the same dimension as same.

[0045] It should be noted that the purpose of sparse projection is to reduce gradient communication while decreasing the number of dimensions that need to be added for differential privacy noise. Projecting to a low-dimensional sparse subspace means adding Laplacian noise only to non-zero gradient components, thereby achieving a smaller noise amplitude under the same privacy budget and improving the effectiveness of federated gradient aggregation.

[0046] Step 5: Cross-organization node alignment and global embedding fusion.

[0047] At the federated aggregation node, cross-institutional node alignment is performed on the privacy-preserved dual-channel local node embedding vector sets uploaded by each institution based on anonymous node encoding. For the embedding vectors of the same anonymous node in different institutions, they are weighted and concatenated according to the neighborhood-scale normalization coefficient to output a global dual-channel fused node embedding matrix.

[0048] Step 5 may specifically include: Step 501: Collect the privacy-protected dual-channel local node embedding vector sets uploaded by various institutions at the federated aggregation node, and identify the corresponding embedding vectors of the same counterparty in different institutions based on the anonymous node encoding.

[0049] Step 502: For the same anonymous node, encode multiple embedding vectors and calculate the neighborhood size normalization coefficient of that node in the source organization of each embedding vector. Let the anonymous node be... In the The neighborhood size in each institution is That is, the node in the local graph of the organization. The number of neighboring nodes, For containing anonymous nodes The total number of institutions, then the neighborhood size normalization coefficient The calculation formula is: ; in, For anonymous node indexing, For institutional indexing, Anonymous node In the The number of neighboring nodes in each institution For containing anonymous nodes The total number of institutions, The index for traversing the structure in the summation.

[0050] Step 503, the same anonymous node Privacy-protected dual-channel local node embedding vectors in various institutions Weighted by neighborhood size normalization coefficients and then concatenated: ; in, Anonymous node The global dual-channel fusion node embedding vector, Anonymous node In the Privacy-protected dual-channel local node embedding vectors in each institution This is the corresponding neighborhood size normalization coefficient. This represents a multi-vector concatenation operation. It arranges the global dual-channel fused node embedding vectors of all anonymous nodes in rows, generating a global dual-channel fused node embedding matrix.

[0051] In this embodiment, for anonymous nodes appearing only in a single institution, their neighborhood size normalization coefficient is 1, and the global dual-channel fused node embedding vector is the privacy-preserved dual-channel local node embedding vector within that institution. For anonymous nodes appearing in multiple institutions, the global embedding dimension generated by splicing together the data from each institution is different from the number of institutions. Relatedly, in subsequent steps, the embedded sub-vectors are extracted by channel dimension rather than by institution dimension. That is, the data difference channel sub-vector and the time delay channel sub-vector are respectively composed of the weighted concatenation of the corresponding channels of each institution.

[0052] Step 6: Generate the joint anomaly weighted association graph.

[0053] For each pair of nodes with a transaction relationship in the global dual-channel fusion node embedding matrix, embedding sub-vectors are extracted from the data difference channel and the time delay channel, respectively. Field deviation metric and delay collaboration metric are calculated and mapped to joint anomaly edge weights through a learnable nonlinear fusion function to generate a joint anomaly weighted association graph.

[0054] Step 6 may specifically include: Step 601: For each pair of nodes with a transaction relationship in the global dual-channel fusion node embedding matrix... Extract nodes respectively and nodes Subvectors belonging to the data difference channel in the global dual-channel fusion node embedding vector and Calculate the field deviation metric: ; in, and This represents the indices of the two nodes in the node pair. and They are nodes and nodes Data difference channel subvectors, This is an L2 norm operation.

[0055] Step 602, extract nodes respectively and nodes Subvectors belonging to the timing delay channel in the global dual-channel fusion node embedding vector and Calculate the delay co-operation metric: ; in, and They are nodes and nodes The timing delay channel subvector.

[0056] It should be noted that the dimensional consistency of the two sub-vectors involved in the calculations in steps 601 and 602 is guaranteed as follows: In step 5, the embedding vectors of the same anonymous node in each institution are weighted and concatenated according to the institution dimension. For node pairs... When extracting the data difference channel subvectors, the data difference channel components corresponding to each institution are extracted and concatenated in the same institutional order. and All dimensions are ,in The number of organizations participating in the overall assembly. For the first The embedding dimension of data difference channels within an organization. When a node and nodes When the sets of mechanisms are not completely identical, zero vectors are used to fill the corresponding channel embedding positions of the missing mechanisms to ensure that the two sub-vectors have the same dimension. (Temporal delay channel sub-vector) and The dimension alignment is the same as the data difference channel, and both have the same dimension. ,in For the first The embedding dimension of the timing delay channel in each mechanism is used, and the corresponding positions of missing mechanisms are also filled with zero vectors.

[0057] Step 603, the field deviation measure value and delay co-measures The input is a learnable nonlinear fusion function, mapped to joint anomaly edge weights: ; in, For parameters A multilayer perceptron, with a two-dimensional vector as input. The output layer is a single-neuron fully connected layer connected to a sigmoid activation function, and the output is a scalar. Multilayer perceptron The algorithm takes the field deviation metric and latency collaboration metric of all node pairs with transactional relationships in the global dual-channel fusion node embedding matrix as input, and uses the label indicating whether the corresponding node pair belongs to a known anomalous association as the supervision signal. It is trained using the binary cross-entropy loss function and updated with the Adam optimization algorithm. Assign the weights of the joint anomaly edges of all node pairs to the corresponding transaction relationship edges to generate a joint anomaly weighted association graph.

[0058] In the embodiments of this application, the nonlinear fusion function This results in a superlinear enhancement of the joint anomaly edge weights when both the field deviation metric and the delay co-occurrence metric are high. Specifically, when both types of anomaly signals co-occur, the joint anomaly edge weights are greater than the sum of the weights of the two signals occurring individually. This superlinear characteristic allows weak signals scattered across the data difference and temporal delay dimensions in collaborative fraud scenarios to be effectively amplified after fusion.

[0059] Step 7: Identification of abnormal paths based on hop-by-hop cumulative propagation.

[0060] Perform hop-by-hop cumulative propagation calculation on the joint anomaly weighted association graph, accumulate joint anomaly edge weights along connected paths, identify the longest connected path whose cumulative joint anomaly score exceeds an adaptive threshold, and generate a set of dual-channel joint differential propagation paths and the cumulative joint anomaly score of each path.

[0061] Step 7 may specifically include: Step 701: On the joint anomaly weighted association graph, perform a hop-by-hop path search for each node as the starting node. Accumulate the joint anomaly edge weights along each connected path to obtain the cumulative joint anomaly score for that path. Assume a path passes through nodes sequentially. Then the cumulative joint anomaly score of this path The calculation formula is: ; in, This represents the number of nodes contained in the path. This represents the traversal index of the edges in the path. and The first in the path The start and end nodes of the edge. For nodes With nodes The joint abnormal edge weights between them.

[0062] Step 702: Determine the adaptive threshold based on the historical normal association group embedding distribution. Collect the cumulative distribution of joint abnormal edge weights corresponding to association groups that were not marked as abnormal in historical audits, and calculate the mean of this distribution. and standard deviation Adaptive threshold The calculation formula is: ; in, This represents the mean of the cumulative distribution of the joint abnormal edge weights of historical normal association groups. Let be the standard deviation of this distribution. This is an adjustable threshold sensitivity parameter.

[0063] Step 703: For all cumulative joint anomaly scores exceeding the adaptive threshold The longest connected path is identified among the paths. When multiple paths have sub-path inclusion relationships, the longest path is retained and its sub-paths are removed. The filtered paths are used to form a set of dual-channel joint differential propagation paths, and the cumulative joint anomaly score of each path is recorded.

[0064] In this embodiment of the application, to reduce the computational complexity of hop-by-hop path search, step 701 can adopt the following pruning strategy: During the hop-by-hop search process, when the estimated value of the accumulated anomaly score from the starting node to the current node plus the weight of the largest joint anomaly edge among the remaining reachable nodes multiplied by the remaining maximum number of hops is still less than the adaptive threshold. If this happens, the search for that path will be terminated.

[0065] Step 8: Generate cross-agency confirmation data privacy processing results.

[0066] The anonymous nodes on each path in the dual-channel joint difference propagation path set are mapped back to the corresponding institutional confirmation records. The main deviation fields of the data difference channel and the delay pattern characteristics of the time-series delay channel are labeled for each path. A cross-institutional confirmation data privacy processing result report is generated, sorted in descending order by cumulative joint anomaly score.

[0067] Step 8 may specifically include: Step 801: For each path in the dual-channel joint differential propagation path set, query the correspondence between the anonymous node codes and confirmation records stored locally by each institution according to the anonymous node codes on the path, and map the anonymous nodes back to the corresponding institution confirmation records.

[0068] Step 802: For the global dual-channel fusion node embedding vector of adjacent node pairs in each path, extract the differences in each dimension of the data difference channel sub-vector, identify the original transaction field corresponding to the dimension with the largest absolute value of the difference, and label it as the main deviation field. Extract the distribution of the differences in each dimension of the time-series delay channel sub-vector and label the delay mode features.

[0069] Step 803: Sort all paths from highest to lowest cumulative joint anomaly score, and generate a cross-agency confirmation data privacy processing result report. Each record in the cross-agency confirmation data privacy processing result report includes the confirmation records of the agencies involved in the path, the path length, the cumulative joint anomaly score, the main deviation fields of each side, and the delay pattern characteristics.

[0070] On the other hand, this invention also proposes a cross-institutional confirmation data privacy processing system based on federated learning, including: The local feature extraction and anonymization module is used to obtain the confirmation response data stored locally by each participating institution, extract the edge feature vector and time delay feature and concatenate them into a dual-channel edge feature vector, anonymize the counterparty identifier and the auditee identifier to generate anonymous node codes and anonymous association identifiers, and assemble the local transaction relationship adjacency fragments.

[0071] The cross-institutional association topology generation module is used to receive anonymous association identifiers uploaded by various institutions, perform privacy set intersection operations to identify the set of anonymous association identifiers shared by the institutions, generate a cross-institutional confirmation association topology index, and distribute it back to the relevant institutions.

[0072] The local dual-channel graph neural network processing module is used to aggregate association groups based on the cross-institutional confirmation association topology index, input the local transaction relationship adjacency fragments and dual-channel edge feature vectors into the local dual-channel graph neural network, process the corresponding edge feature sub-vectors through the data difference message passing channel and the time delay message passing channel respectively, and perform feature cross-enhancement through cross-channel attention gating, and output dual-channel local node embedding vectors.

[0073] The differential privacy protection and federated aggregation module is used to apply adaptive differential privacy noise injection based on node degree to the dual-channel local node embedding vector, perform sparse projection processing on the training gradient and add calibrated Laplacian noise, perform safe aggregation of gradients on the federated aggregation node and distribute the global aggregated gradient, and output the trained local dual-channel graph neural network and the privacy-protected dual-channel local node embedding vector set after multiple rounds of iteration.

[0074] The cross-institutional node alignment and fusion module is used to perform cross-institutional node alignment on the privacy-protected dual-channel local node embedding vector set uploaded by each institution based on anonymous node encoding, and generate a global dual-channel fused node embedding matrix by weighted concatenation according to the neighborhood scale normalization coefficient.

[0075] The joint anomaly weighted association graph generation module is used to calculate the field deviation metric and delay collaboration metric for each pair of nodes with transaction relationships in the global dual-channel fusion node embedding matrix, and map them to the joint anomaly edge weights through a learnable nonlinear fusion function to generate a joint anomaly weighted association graph.

[0076] The anomaly path identification module is used to perform hop-by-hop cumulative propagation calculation on the joint anomaly weighted association graph, accumulate joint anomaly edge weights along connected paths, identify the longest connected path whose cumulative joint anomaly score exceeds an adaptive threshold, and generate a set of dual-channel joint differential propagation paths and the cumulative joint anomaly score of each path.

[0077] The results report generation module is used to map anonymous nodes on each path in the dual-channel joint differential propagation path set back to the corresponding institutional confirmation records, annotate the main deviation fields and delay mode features, sort them in descending order by cumulative joint anomaly score, and generate a cross-institutional confirmation data privacy processing results report.

[0078] By way of example, the system proposed in this invention may also include various features and combinations thereof in the method embodiments, which will not be elaborated here.

[0079] The method of this invention anonymizes the counterparty identifier and the auditee identifier using a one-way hash function and a locality-sensitive hash function, respectively. It also uses privacy set intersection operation to establish a cross-institutional association topology index at the federated coordination node. Each institution does not need to share plaintext identifiers or complete transaction relationship graphs. Therefore, the transaction topology of related parties and the information on the attribution of audit tasks are not exposed during the federated learning process, thus solving the privacy leakage problem in the cross-institutional association relationship establishment stage.

[0080] By performing message passing computation of the local dual-channel graph neural network locally in each institution, and applying adaptive differential privacy noise based on node degree to the local node embedding vector of the dual-channel node, and performing sparse projection differential privacy processing on the gradient before uploading it to the federated aggregation node, the transaction relationship structure and association group statistical features do not leave the local institution. Therefore, the original data and graph structure information of each institution are protected during the federated aggregation and gradient passing process.

[0081] Cross-channel attention gating enables feature cross-enhancement of data difference signals and time-delay signals on a shared topology. The nonlinear fusion function enables superlinear enhancement when the two types of signals co-occur. Therefore, weak abnormal signals scattered in multi-hop transaction chains and multi-institution response sequences in collaborative fraud scenarios are effectively amplified after fusion and hop-by-hop cumulative propagation, solving the problem that single-dimensional processing cannot effectively identify collaborative anomalies due to weak signals.

[0082] See Figures 2-9 The following application examples are proposed: A corporate group consists of a parent company and three subsidiaries, denoted as Institution A, Institution B, Institution C, and Institution D. The auditor initiated a cross-institutional confirmation audit of the group, focusing on verifying the consistency of accounts receivable and payable data among the various institutions. Institution A and Institution B have numerous supply chain settlement transactions; Institution C and Institution D have service fee transactions; and Institution B and Institution C have cross-subsidiary fund transfer records. Prior to the audit, each institution had stored its own confirmation response data locally, but was not allowed to directly disclose the original transaction records and counterparty identification information to other institutions or the auditor. The core concern of this audit is whether Institution B and Institution C colluded to adjust accounts receivable and payable amounts and simultaneously delayed confirmation responses.

[0083] In step 1, each institution extracts the edge feature vector and time-series delay feature of each transaction record from the local confirmation response data. Taking three local confirmation records of Institution B as an example, the transaction amount range code and transaction type code are converted by one-hot encoding, the transaction frequency is standardized by Z-score, and the response time interval is standardized by Z-score and then encoded as a time-series delay feature. Subsequently, the counterparty identifier is mapped to an anonymous node code using a one-way hash function, and the auditee identifier is encoded as an anonymous association identifier using locality-sensitive hashing. These are then assembled together with the dual-channel edge feature vector to form a local transaction relationship adjacency fragment.

[0084] Table 1. Feature Extraction and Anonymization Results of Local Confirmation Response Data from Institution B

[0085] In step 2, each institution uploads its locally generated anonymous association identifiers to the federated coordination node. The federated coordination node performs a privacy set intersection operation on the sets of anonymous association identifiers uploaded by institutions A, B, C, and D, identifies anonymous association identifiers that have confirmation records in multiple institutions simultaneously, and distributes the matching results back to the relevant institutions in the form of association indexes.

[0086] Table 2. Cross-institutional privacy set intersection results and associated topology index

[0087] In step 3, each institution aggregates local confirmation response features belonging to the same anonymous association identifier into association groups based on the cross-institutional confirmation association topology index. Taking institution B as an example, the association group corresponding to the anonymous association identifier LSH-C7 contains two records, REC-E01 and REC-E02. After the local transaction relationship adjacency fragments and dual-channel edge feature vectors are input into the local dual-channel graph neural network, the data difference message passing channel processes the edge feature vector sub-vectors, and the time delay message passing channel processes the time delay feature vector sub-vectors; both share the same adjacency structure. After each layer of graph convolution is completed, the interaction attention weights of the cross-channel attention gating computation node H3A9F2 are calculated. Feature cross-enhancement is performed on the node embeddings of both channels. Since node H3A9F2 exhibits high normalized temporal delay values ​​in both REC-E01 and REC-E02, its... As the value approaches a higher level, the time delay anomaly features are enhanced and propagated to the data difference channel. After multi-layer processing, the output is a dual-channel local node embedding vector containing both the data difference channel component and the time delay channel component.

[0088] In step 4, adaptive differential privacy noise based on node degree is applied to the dual-channel local node embedding vectors of each node in the local map of Organization B. Node H3A9F2 has a degree of 5 in this map, node B5D1E8 has a degree of 2, and the maximum degree in this local map is... The basic noise scale parameter is 6. After configuration, the two nodes obtain superimposed Laplacian noise of different amplitudes. Simultaneously, sparse projection processing is performed on the gradients of this training round, retaining the components with the highest absolute values ​​and adding calibrated Laplacian noise before uploading them to the federated aggregation node.

[0089] Table 3. Degree of Node B and Noise Scaling Factor

[0090] In step 5, the federated aggregation node gathers the privacy-preserved dual-channel local node embedding vector sets uploaded by Institution B and Institution C, and identifies the embedding vector of the node in both institutions based on the anonymous node code H3A9F2. The neighborhood size of node H3A9F2 in Institution B is 5, and the neighborhood size of the same node in Institution C is 3. The neighborhood size normalization coefficient is 2. After separate calculations, the privacy-preserving embedding vectors of the two institutions are weighted and concatenated to generate the global dual-channel fusion node embedding vector of that node, which is then arranged together with the results of all other nodes to form a global dual-channel fusion node embedding matrix.

[0091] Table 4 Cross-mechanism embedding and fusion process of node H3A9F2

[0092] In step 6, for each pair of nodes with a transaction relationship in the global dual-channel fusion node embedding matrix, the data difference channel sub-vector and the time delay channel sub-vector are extracted respectively, and the field deviation metric is calculated. and delay co-measures Input a learnable nonlinear fusion function The mapping is represented by joint anomaly edge weights. Taking the node pair (H3A9F2, B5D1E8) as an example, both its data difference channel and time delay channel exhibit high deviations. The nonlinear fusion function produces a superlinear enhancement effect, and the joint anomaly edge weights are significantly higher than those corresponding to a single channel.

[0093] Table 5. Calculation results of joint anomaly edge weights for typical node pairs

[0094] In step 7, a hop-by-hop path search is performed using each node in the joint anomaly weighted association graph as the starting node, accumulating the joint anomaly edge weights along the connected paths. The mean is calculated based on the cumulative distribution of the joint anomaly edge weights of historical normal association groups. and standard deviation Set threshold sensitivity parameters Then an adaptive threshold is obtained. The two-hop path containing nodes H3A9F2—B5D1E8 has the highest cumulative joint anomaly score, exceeding [a certain threshold]. It was identified as the longest connectivity anomaly path and included in the dual-channel joint differential propagation path set.

[0095] Table 6. Cumulative scores and identification results for major abnormal paths.

[0096] In step 8, for path PATH-01 in the dual-channel joint difference propagation path set, the anonymous nodes on the path are encoded and mapped back to the specific confirmation records of Institution B and Institution C (REC-E01, REC-E02, and the corresponding records of Institution C). The differences in each dimension of the data difference channel sub-vectors are extracted, and the "transaction amount range" is identified as the main deviation field. The distribution of the time-series delay channel sub-vector differences is extracted and labeled as "centralized delay mode." The results are sorted in descending order by the cumulative joint anomaly score, generating a cross-institutional confirmation data privacy processing result report.

[0097] Table 7. Excerpt from the report on data privacy processing results of cross-agency confirmations.

[0098] Throughout the process, the data starts from the original confirmation response records of Institutions B and C. Step 1 involves feature extraction and anonymization, converting the data into dual-channel edge feature vectors and anonymous identifiers. Step 2 uses privacy set intersection to confirm LSH-C7 as a cross-institutional shared association identifier. Step 3 uses local dual-channel graph neural network processing to generate node embeddings that fuse data differences and temporal delay signals. Step 4 uses adaptive differential privacy processing to protect the original graph structure information. Step 5 uses cross-institutional node alignment and weighted concatenation with neighborhood-scale normalized coefficients to generate global embeddings. Step 6 uses a nonlinear fusion function to transform the co-occurrence effect of two types of anomalous signals into joint anomalous edge weights. Step 7 uses hop-by-hop cumulative propagation to amplify the collaborative fraud signal along the two-hop path. Finally, Step 8 maps the data to traceable confirmation records and generates a report. The output data of each step serves as the direct input to the next step, forming a complete and logically consistent data flow chain.

[0099] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for handling cross-institutional confirmation data privacy based on federated learning, characterized in that, include: Each participating institution extracts edge feature vectors and time delay features from local confirmation response data and concatenates them into dual-channel edge feature vectors. It then anonymizes the counterparty identifier and the auditee identifier to generate anonymous node codes and anonymous association identifiers, and assembles local transaction relationship adjacency fragments. The federal coordination node performs a privacy set intersection operation on the anonymous association identifiers uploaded by each agency to generate a cross-agency confirmation association topology index; Each institution inputs the local transaction relationship adjacency fragment and the dual-channel edge feature vector into the local dual-channel graph neural network, performs graph convolution operations through the data difference message passing channel and the time delay message passing channel respectively, performs feature cross-enhancement through cross-channel attention gating, and outputs the dual-channel local node embedding vector; An adaptive differential privacy noise injection based on node degree is applied to the dual-channel local node embedding vector. After sparse projection processing of the training gradient, calibrated Laplacian noise is added. The gradient is safely aggregated through federated aggregation nodes and iteratively trained until convergence. Based on the privacy protection of each institution through anonymous node encoding, the dual-channel local node embedding vectors are aligned across institutions, and then weighted and concatenated according to the neighborhood scale normalization coefficient to generate a global dual-channel fused node embedding matrix. For each node pair, calculate the field deviation metric and the delay collaboration metric, and map them to the joint anomaly edge weights through a nonlinear fusion function to generate a joint anomaly weighted association graph. Perform hop-by-hop cumulative propagation on the joint anomaly weighted association graph to identify paths where the cumulative joint anomaly score exceeds an adaptive threshold, and generate privacy-preserving results for cross-agency confirmation data.

2. The cross-institutional confirmation data privacy processing method based on federated learning according to claim 1, characterized in that, The edge feature vector includes transaction amount range encoding, transaction frequency, and transaction type encoding. The transaction amount range encoding and transaction type encoding are converted into numerical representations using one-hot encoding, and the transaction frequency is processed using Z-score normalization. The time delay feature is generated by encoding the response time interval of the confirmation after Z-score normalization. The anonymous node encoding is generated by mapping the counterparty identifier using a one-way hash function, and the anonymous association identifier is generated by encoding the audited entity identifier using locality-sensitive hashing.

3. The cross-institutional confirmation data privacy processing method based on federated learning according to claim 1, characterized in that, The calculation process for the cross-channel attention gating includes: Let the data difference channel be the first... Layer output nodes The embedding is The timing delay channel is in the first Layer output nodes The embedding is ,in This refers to the layer index of the graph convolutional layer. For node indexing; the interaction attention weights are: ； in This is a learnable attention weight vector. For bias scalars, It is the sigmoid activation function. This is a vector concatenation operation; feature cross enhancement is as follows: ； ； in and These are the enhanced data difference channel node embedding and the timing delay channel node embedding, respectively.

4. The cross-institutional confirmation data privacy processing method based on federated learning according to claim 3, characterized in that, The cross-channel attention gating adopts a multi-head attention approach, which will... and Each part is equally divided along its feature dimension into Subvectors, of which Given the number of attention heads, calculate independent attention weights for each head and perform feature cross-enhancement, then concatenate the outputs of each head to restore the original structure.

5. The cross-institutional confirmation data privacy processing method based on federated learning according to claim 1, characterized in that, The adaptive differential privacy noise injection based on node degree includes: Set nodes The degree in the adjacency segment of the local transaction relation is , This represents the maximum degree of all nodes in the local graph. Based on the noise scale parameter, noise scaling factor for: ； in Indexing nodes; indexing nodes Dual-channel local node embedding vector Embedding with added noise ,in From Noise vectors sampled independently from the distribution; The sparse projection processing includes preserving the gradient with the absolute value listed first. The proportion is set to one component and the remaining components are set to zero, where For the sparsity parameter, calibrated Laplacian noise is added to the retained non-zero gradient components.

6. The cross-institutional confirmation data privacy processing method based on federated learning according to claim 1, characterized in that, The weighted concatenation based on neighborhood size normalization coefficient includes: Set anonymous nodes In the The number of neighbor nodes in each institution is , For containing anonymous nodes The total number of institutions, and the neighborhood size normalization coefficient are: ； in For anonymous node indexing, For institutional indexing, Traversing the index for the organization in the summation; anonymous nodes The global dual-channel fusion node embedding vector is: ； in Anonymous node In the The privacy-preserved dual-channel local node embedding vector in each institution; for anonymous nodes that appear only in a single institution, their global dual-channel fused node embedding vector is the corresponding privacy-preserved dual-channel local node embedding vector in that institution.

7. The cross-institutional confirmation data privacy processing method based on federated learning according to claim 1, characterized in that, The field deviation metric and the delay co-operation metric are the L2 norm distances between node pairs in the data difference channel sub-vectors and the time delay channel sub-vectors, respectively; the nonlinear fusion function is a multilayer perceptron, with the input being a two-dimensional vector composed of the field deviation metric and the delay co-operation metric, and the output layer connected to a sigmoid activation function to output the joint anomaly edge weights. When the sets of mechanisms appearing in the two nodes of a node pair are not completely identical, zero vectors are filled into the corresponding channel embedding positions of the missing mechanisms.

8. The cross-institutional confirmation data privacy processing method based on federated learning according to claim 1, characterized in that, The hop-by-hop cumulative propagation includes: Suppose a path passes through nodes in sequence. The cumulative joint anomaly score for this path is: ； in This represents the number of nodes contained in the path. This represents the traversal index of the edges in the path. The joint outlier edge weights between adjacent nodes; The adaptive threshold is determined based on the cumulative distribution of the joint abnormal edge weights of historical normal association groups: ； in Let be the mean of this distribution. Standard deviation This is a threshold sensitivity parameter; for cumulative joint anomaly scores exceeding... When multiple paths have sub-path inclusion relationships, retain the longest path and remove its sub-paths.

9. The cross-institutional confirmation data privacy processing method based on federated learning according to claim 8, characterized in that, The hop-by-hop cumulative propagation employs a pruning strategy: during the hop-by-hop search process, if the accumulated anomaly score from the starting node to the current node plus the estimated value of the maximum joint anomaly edge weight among the remaining reachable nodes multiplied by the remaining maximum hop count is less than an adaptive threshold... If this happens, the search for that path will be terminated.

10. A cross-institutional confirmation data privacy processing system based on federated learning, used to execute the cross-institutional confirmation data privacy processing method based on federated learning as described in any one of claims 1 to 9, characterized in that, include: The local feature extraction and anonymization module is used to extract edge feature vectors and time delay features from local confirmation response data and concatenate them into dual-channel edge feature vectors. It anonymizes the counterparty identifier and the auditee identifier to generate anonymous node codes and anonymous association identifiers, and assembles local transaction relationship adjacency fragments. The cross-institutional association topology generation module is used to receive anonymous association identifiers uploaded by various institutions, perform privacy set intersection operation to identify the set of anonymous association identifiers shared by the institutions, generate a cross-institutional confirmation association topology index and distribute it back to the relevant institutions. The local dual-channel graph neural network processing module is used to aggregate association groups based on the cross-institutional confirmation association topology index, input the local transaction relationship adjacency fragments and dual-channel edge feature vectors into the local dual-channel graph neural network, process the corresponding edge feature sub-vectors through the data difference message passing channel and the time delay message passing channel respectively, perform feature cross-enhancement through cross-channel attention gating, and output dual-channel local node embedding vectors. The differential privacy protection and federated aggregation module is used to apply adaptive differential privacy noise injection based on node degree to the dual-channel local node embedding vector, perform sparse projection processing on the training gradient and add calibrated Laplacian noise, perform safe aggregation of gradients on federated aggregation nodes and distribute global aggregated gradients, and output the trained local dual-channel graph neural network and the privacy-protected dual-channel local node embedding vector set after multiple rounds of iteration. The cross-institutional node alignment and fusion module is used to perform cross-institutional node alignment on the privacy-protected dual-channel local node embedding vector set uploaded by each institution based on anonymous node encoding, and generate a global dual-channel fused node embedding matrix by weighted concatenation according to the neighborhood scale normalization coefficient. The joint anomaly weighted association graph generation module is used to calculate the field deviation metric and the delay collaboration metric for each pair of nodes with transaction relationships in the global dual-channel fusion node embedding matrix, and map them to the joint anomaly edge weights through a nonlinear fusion function to generate a joint anomaly weighted association graph. The abnormal path identification module is used to perform hop-by-hop cumulative propagation calculation on the joint anomaly weighted association graph, accumulate joint anomaly edge weights along connected paths, identify the longest connected path whose cumulative joint anomaly score exceeds an adaptive threshold, and generate a set of dual-channel joint differential propagation paths and the cumulative joint anomaly score of each path. The results report generation module is used to map anonymous nodes on each path in the dual-channel joint differential propagation path set back to the corresponding institutional confirmation records, annotate the main deviation fields and delay mode features, sort them in descending order by cumulative joint anomaly score, and generate a cross-institutional confirmation data privacy processing results report.

Images