An inter-agency financial data federated learning modeling system and privacy compliance verification method

Abstract

The application discloses a cross-agency financial data federated learning modeling system and a privacy compliance verification method, relates to the technical field of financial data processing, and comprises the following: a federated aggregation module, which is used for aligning financial time sequence characteristics of each agency by a timestamp hashing alignment strategy on a server side, and aggregating the aligned financial time sequence characteristics by using a federated time convolution network to generate time sequence model parameters for capturing cross-agency global time sequence dependence; a federated reasoning module, which is used for generating a minimum spanning tree representation of global enterprise correlation relationship based on a local subgraph of enterprise correlation relationship constructed by each agency client on the server side based on a secure multi-party computing strategy; the server side aggregates graph embedding vectors by using a federated graph attention network and performs cross-agency graph reasoning in combination with the minimum spanning tree representation to construct graph model parameters for identifying enterprise correlation risks; and a model joint training module, which is used for fusing the time sequence model parameters and the graph model parameters on the server side to determine global model parameters.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial data processing, in particular to a cross-institution financial data federated learning modeling system and a privacy compliance verification method. BACKGROUND

[0002] In the cross-institution data collaborative analysis scenario in the financial field, the processing of multi-modal data is crucial for risk identification and model construction. However, the existing technology has significant deficiencies in data compliance verification and privacy protection.

[0003] For time series data processing, traditional methods lack deep integration of business cycle characteristics, cannot effectively verify the integrity and rationality of time interval division, and are prone to missing key data segments or mixing unauthorized range data. At the same time, with the increasing requirements of financial supervision on data use compliance, how to achieve privacy protection and fine-grained compliance verification in the whole process in cross-institution data collaboration has become a technical problem to be solved.

[0004] In view of the above problems, no effective solution has been proposed so far. SUMMARY

[0005] The embodiments of the present application provide a cross-institution financial data federated learning modeling system and a privacy compliance verification method to solve the above technical problems.

[0006] The present application provides a cross-institution financial data federated learning modeling system, comprising:

[0007] A federated aggregation module is configured to extract financial time series features of financial data from each institution client and attach time interval labels, and send them to a server; the server aligns the financial time series features of each institution through a timestamp hashing alignment strategy, dynamically adjusts the federated aggregation frequency based on the financial business cycle, and uses a federated time convolution network to aggregate the aligned financial time series features to generate time series model parameters that capture cross-institution global time series dependencies;

[0008] A federated inference module is configured to generate a minimum spanning tree representation of global enterprise association relationships based on a secure multi-party computation strategy based on local subgraphs of enterprise association relationships constructed by each institution client; each institution client generates a graph embedding vector based on the local subgraph and sends it to the server; the server uses a federated graph attention network to aggregate the graph embedding vectors and combines the minimum spanning tree representation for cross-institution graph inference to construct graph model parameters for identifying enterprise association risks;

[0009] A model joint training module is configured to fuse the time series model parameters and the graph model parameters to determine global model parameters.

[0010] Further, when the federal aggregation module performs financial time series feature aggregation, it includes:

[0011] The server receives the time interval labels attached by each institution client, maps the financial business cycle to a time encoding vector, where the financial report disclosure quarter corresponds to a high-dimensional activation vector, and the regular cycle corresponds to a benchmark vector;

[0012] The attention calculation unit of the federal time convolution network integrates the time encoding vector and dynamically adjusts the aggregation weight by identifying the time series feature fluctuation pattern of the same industry institutions, generating a synergistic enhancement signal for the institution group whose fluctuation phase synchronization degree exceeds the industry threshold; during the financial report disclosure quarter, activating the dedicated feature extraction channel to strengthen the convolution kernel response intensity of the quarterly profit table data; constructing a bidirectional attention link for cross-institutional features that have a dependent relationship, forcing the bank loan data and associated merchant repayment data to be associated during the aggregation process;

[0013] Based on the current business cycle state, dynamically select the output layer channel, when detecting seasonal risk patterns, enable the long cycle convolution kernel group to output capture cross-year dependent time series model parameters.

[0014] Further, the cross-institutional graph reasoning in the federal inference module includes:

[0015] The server extracts the guarantee ring topology and supply chain hierarchical relationship from the minimum spanning tree representation to generate a risk transmission path matrix;

[0016] The federal graph attention network performs neighbor aggregation by performing the following operations: impose a lower limit constraint on the attention coefficient for nodes within the guarantee ring, so that their mutual attention is not lower than a preset business safety threshold; calculate the hierarchical attenuation factor based on the risk transmission path matrix, and impose a weighted penalty on the risk output of the core enterprise node, which is positively related to the number of its upstream and downstream connections; build cross-institutional virtual edges for nodes with multiple associations to simulate implicit guarantee relationships in the attention layer; inject adversarial test samples to simulate guarantee chain breakage scenarios and verify risk transmission early warning capabilities, and dynamically adjust the graph regularization strength based on the verification results.

[0017] Further, the construction of the bidirectional attention link includes:

[0018] The server analyzes the time lag correlation between the aligned bank loan data and merchant repayment data;

[0019] Mark the institution pair with a stable time lag within a preset interval as a strong dependent group, and generate a feature correlation certificate between institutions;

[0020] Deploy a certificate verification unit in the aggregation layer of the federal time convolution network, allowing only features from certified institution groups to participate in bidirectional attention calculation.

[0021] Further, the calculation of the hierarchical attenuation factor comprises:

[0022] Analyzing the transaction frequency and default conduction history data in the supply chain hierarchical relationship;

[0023] Setting a basic attenuation coefficient for the direct upstream and downstream nodes, and starting the conduction strengthening mode when the transaction frequency exceeds the industry average;

[0024] Using an exponential attenuation strategy for cross-hierarchical conduction paths, and increasing the attenuation strength by a preset multiple for each additional hierarchical level.

[0025] Further, the generation of the adversarial test sample comprises:

[0026] Identifying the core enterprise node of the guarantee ring topology in the minimum spanning tree representation;

[0027] Simulating a sudden default event of the core enterprise node, and diffusing the default state along the risk conduction path matrix layer by layer;

[0028] Generating a timestamped virtual default event sequence injection test data set.

[0029] Further, the control of the periodic adaptive output comprises:

[0030] Establishing a financial seasonality feature library to store historical seasonal risk fluctuation feature patterns;

[0031] Real-time comparison of the similarity between the current time series features and the feature library;

[0032] When the similarity exceeds the matching threshold, automatically switch to the long-period convolution kernel group and freeze the regular output channel.

[0033] The present application provides a privacy compliance verification method for cross-institution financial data federated learning modeling, comprising:

[0034] Based on the financial business cycle rules, verify whether the time interval label attached to the financial time series features uploaded by each institution client covers the complete business cycle core data segment;

[0035] Call the dynamically adjusted federated aggregation frequency parameter to verify whether the real-time aggregation frequency meets the preset frequency range of the current business cycle;

[0036] For the time series model parameters aggregated by the federated time convolution network, detect the sensitive feature leakage risk through differential noise disturbance, and verify the feature extraction interval compliance based on counterfactual reasoning;

[0037] In the process of generating the minimum spanning tree representation by secure multi-party computation, the sensitive field hiding and the correlation disclosure permission of each local subgraph are verified by zero-knowledge proof;

[0038] The dimension of the graph embedding vector generated by the verification agency client matches the complexity of the local subgraph;

[0039] The attention weight distribution of the federated graph attention network is monitored to locate and block illegal association edge reasoning;

[0040] Cross-modal illegal association in the fusion process of timing model parameters and graph model parameters is detected, and the rule coupling degree is reduced by feature orthogonalization.

[0041] Further, in the process of verifying the time interval label based on the financial business cycle rule, the closed-loop repair is realized by dynamic compliance control through the timestamp hash alignment strategy, specifically including:

[0042] Establish the binding rule of financial data types and anti-collision hash algorithm: configure SHA-3 algorithm for credit data, and configure BLAKE3 algorithm for high-frequency transaction data;

[0043] When it is detected that the agency client uses a non-binding algorithm, real-time feature transmission is blocked and an algorithm reset instruction is issued;

[0044] Load the time deviation threshold corresponding to the current business cycle type of the system: allow ±4 hours deviation for daily transaction cycle, and allow ±1 working day deviation for quarterly settlement cycle;

[0045] If the agency timestamp deviation exceeds the time deviation threshold, activate the data repair channel: intercept the original time series data in the deviation period, and generate a compensation feature vector by calling the local aggregation function of the federated time convolution network;

[0046] The compensation feature vector is attached with a time interval label and resubmitted to the verification process, and the cycle is executed until the complete business cycle core data segment coverage detection is passed.

[0047] Further, the method further comprises:

[0048] In the process of secure multi-party computation, a real-time auditor is implanted, and when it is detected that the guarantee relationship edge weight is greater than 40% of the registered capital of the associated enterprise, the over-limit edge is cropped and a violation event snapshot is generated;

[0049] Delete the local subgraph of the agency, and if the global enterprise association risk score change rate is >10%, it is determined that there is a violation of association leakage; freeze the graph embedding vector transmission channel of the agency, and mark it as a high-risk subgraph source;

[0050] Reconstructing the minimum spanning tree representation based on the compliance edge weight threshold: retaining the legal association edges after auditing; injecting virtual balanced nodes into the topological partitions broken by pruning to maintain the connectivity of enterprise risk transmission paths.

[0051] Based on the embodiments provided in the present application, in the application scenario of cross-institutional data collaborative analysis in the financial field, the cross-institutional financial data federated learning modeling system has significant beneficial effects. In the federated aggregation module, each institution client extracts financial time series features and sends them to the server with time interval labels. The server aligns the financial time series features of each institution through a timestamp hashing alignment strategy, solving the alignment problem of cross-institutional financial time series data caused by different time bases, ensuring the consistency and comparability of data in the time dimension of different institutions, and providing an accurate data basis for subsequent model aggregation. Dynamically adjusting the federated aggregation frequency based on the business cycle of finance enables the system to flexibly adjust the data aggregation frequency according to the actual cycle characteristics of financial business, avoiding the problem of missing key data or redundant invalid data caused by fixed frequency aggregation, and improving the efficiency and pertinence of data aggregation. Using the federated time convolution network to aggregate and align the financial time series features, the time series model parameters that capture the global time series dependence of cross-institutional data are generated, which can fully exploit the cross-institutional correlation rules of financial data in the time series, and improve the ability to capture global time series dependence in financial time series data, providing more accurate time series information for financial risk identification and model construction.

[0052] In the federated inference module, the server generates a minimum spanning tree representation of the global enterprise association relationship based on the local subgraphs of enterprise association relationships constructed by each institution client through a secure multi-party computation strategy, realizing the global integration of cross-institutional enterprise association relationships while protecting the sensitive information of local subgraphs, and solving the problem of privacy leakage in traditional methods of integrating cross-institutional enterprise association relationships. Each institution client generates a graph embedding vector and sends it to the server, which uses a federated graph attention network to conduct cross-institutional graph reasoning in combination with the minimum spanning tree representation, constructs graph model parameters for identifying enterprise association risks, and effectively captures the complex association relationship features between enterprises, improving the ability to identify enterprise association risks and providing strong support for financial institutions to accurately assess enterprise association risks in cross-institutional scenarios. The model joint training module fuses the time series model parameters and the graph model parameters to determine the global model parameters, realizes the organic combination of time series information and graph structure information in financial data, fully utilizes the synergistic advantages of multi-modal data, and makes the constructed global model able to comprehensively utilize the time series dependence and enterprise association relationship information of cross-institutional data, more comprehensively and accurately reflecting the inherent characteristics of financial data, and improving the performance of the model in financial risk identification and other tasks.

[0053] The privacy compliance verification method for cross-institution financial data federated learning modeling of the application closely surrounds the financial business cycle and the data privacy protection demand, and realizes fine compliance verification and privacy protection in the cross-institution financial data federated learning process. Based on the financial business cycle rule, it is verified whether the time interval label attached to the financial time series features uploaded by each institution client covers the complete business cycle core data segment, ensuring the integrity of data collection and avoiding the omission of key business cycle data, meeting the requirements of data integrity in data use compliance of financial supervision. The dynamically adjusted federated aggregation frequency parameter is called to verify whether the real-time aggregation frequency meets the preset frequency range of the current business cycle, ensuring that the data aggregation process matches the financial business cycle and avoiding the influence of unreasonable aggregation frequency on data processing effect and compliance, improving the standardization and controllability of the data aggregation process.

[0054] For the time series model parameters aggregated by the federated time convolution network, the sensitive feature leakage risk is detected through differential noise disturbance, and the feature extraction interval compliance is verified based on counterfactual reasoning, which can effectively detect and prevent sensitive feature leakage risk in the time series model parameters, while ensuring that the feature extraction interval meets the specified compliance requirements, fully protecting data privacy and compliance during model training using time series data. In the process of generating the minimum spanning tree representation by secure multi-party computation, the sensitive field hiding and association relationship disclosure permission of each institution's local subgraph are verified by zero-knowledge proof, ensuring that the sensitive fields are effectively hidden when sharing local subgraph information, and the disclosure of the association relationship meets the respective permission regulations, solving the problem of sensitive information protection and permission control in cross-institution data sharing, ensuring the compliance and security of the data sharing process. The dimension of the graph embedding vector generated by the verification institution client matches the complexity of the local subgraph, ensuring that the graph embedding vector can accurately reflect the structure characteristics of the local subgraph, avoiding the problem of data feature loss or redundancy caused by unreasonable dimension, and improving the accuracy and effectiveness of graph data processing. The attention weight distribution of the federated graph attention network is monitored to locate and block illegal association edge reasoning, which can timely discover and prevent the model from learning and utilizing illegal association edges during reasoning, preventing risk misjudgment and data abuse problems caused by illegal association reasoning, and ensuring the compliance and accuracy of the graph reasoning process. Cross-modal illegal association in the fusion process of time series model parameters and graph model parameters is detected, and the feature orthogonalization is used to reduce the violation coupling degree, avoiding illegal association of different modal data in the fusion process, ensuring the compliance of multi-modal data fusion, and improving the legality and reliability of global model parameters, providing comprehensive and effective privacy compliance protection for cross-institution financial data federated learning modeling. BRIEF DESCRIPTION OF DRAWINGS

[0055] The accompanying drawings, which are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0056] Figure 1 A structural diagram of an optional cross-institution financial data federated learning modeling system according to an embodiment of the application;

[0057] Figure 2 A flowchart of cross-institution graph reasoning in an optional federated inference module according to an embodiment of the application.

[0058] Figure 3 A flowchart of a privacy compliance verification method for the optional cross-institution financial data federated learning modeling according to an embodiment of the application.

[0059] The implementation, functional features and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0061] Optionally, as shown in the application provides a cross-institution financial data federated learning modeling system, comprising: Figure 1

[0062] The federated aggregation module 101 is configured to extract financial time series features of financial data of each institution client and attach a time interval label, and send the financial time series features to the server. The server aligns the financial time series features of each institution through a timestamp hashing alignment strategy, dynamically adjusts the federated aggregation frequency based on the financial business cycle, and uses a federated time convolution network to aggregate the aligned financial time series features to generate time series model parameters that capture cross-institution global time series dependencies.

[0063] The federated inference module 102 is configured to generate a minimum spanning tree representation of global enterprise correlation relationships based on the local subgraphs of enterprise correlation relationships constructed by each institution client based on a secure multi-party computation strategy. Each institution client generates a graph embedding vector based on the local subgraph and sends the graph embedding vector to the server. The server uses a federated graph attention network to aggregate the graph embedding vectors and performs cross-institution graph reasoning in combination with the minimum spanning tree representation to construct graph model parameters for identifying enterprise correlation risks.

[0064] ​In this embodiment, the enterprise correlation relationship refers to the business connection between enterprises based on self-owned data records of a single financial institution (such as a bank or a securities company), such as a guarantee relationship, a stock ownership penetration relationship, a supply chain transaction relationship, and the like, to form a local subgraph structure (such as a guarantee chain local graph of a certain bank customer enterprise) containing only the perspective of the institution.

[0065] The global enterprise correlation relationship refers to an integrated abstraction of the enterprise correlation relationship across all participating institutions, which cooperates with each institution local subgraph through secure multi-party computation (such as private set intersection, homomorphic encryption) to generate global topological features (such as a minimum spanning tree representation) of cross-institution enterprise correlation without exposing the original data, reflecting the overall picture of the correlation network of enterprises in multiple-institution dimensions (such as the cross-node of a certain enterprise in the guarantee chain of multiple banks).

[0066] The global enterprise correlation relationship is a topological feature fusion of local enterprise correlation relationships of each institution, which realizes lossless abstraction of cross-institution correlation relationships through privacy protection technology (only statistical features such as degree and connectivity are retained, and no specific enterprise data is involved); the local subgraph is the basic unit for building the global correlation relationship, and the global minimum spanning tree representation is used to capture the key path of the cross-institution correlation risk (such as the hub enterprise in the guarantee chain).

[0067] The model joint training module 103 is configured to fuse the time series model parameters and the graph model parameters by the server to determine the global model parameters.

[0068] Based on the embodiments provided in the present application, in the application scenario of cross-institution data collaborative analysis in the financial field, the cross-institution financial data federated learning modeling system has significant beneficial effects. In the federated aggregation module, each institution client extracts financial time series features and sends them to the server with time interval labels, and the server aligns the financial time series features of each institution through a timestamp hashing alignment strategy, solving the alignment problem of cross-institution financial time series data caused by different time bases, ensuring the consistency and comparability of data in the time dimension of different institutions, and providing an accurate data basis for subsequent model aggregation. Dynamically adjusting the federated aggregation frequency based on the business cycle of finance enables the system to flexibly adjust the data aggregation frequency according to the actual cycle characteristics of financial business, avoiding the problem of missing key data or redundant invalid data caused by fixed frequency aggregation, and improving the efficiency and pertinence of data aggregation. Using the federated time convolution network to aggregate and align the financial time series features, the time series model parameters that capture the cross-institution global time series dependence are generated, which can fully exploit the cross-institution correlation rules of financial data in the time series, and improve the ability of the model to capture global time series dependence in financial time series data, providing more accurate time series information for financial risk identification and model construction.

[0069] In the federated reasoning module, the server generates a minimum spanning tree representation of the global enterprise association relationship based on the local sub-graphs of enterprise association relationship constructed by each agency client through secure multi-party computation strategy. Under the premise of protecting the sensitive information of each agency's local sub-graph, the global integration of cross-agency enterprise association relationship is realized, and the problem of privacy leakage in traditional methods of cross-agency enterprise association relationship integration is solved. Each agency client generates a graph embedding vector and sends it to the server. The server uses federated graph attention network to combine the minimum spanning tree representation for cross-agency graph reasoning, and constructs a graph model parameter for identifying enterprise association risks. It can effectively capture the complex association relationship characteristics between enterprises and improve the identification ability of enterprise association risks, providing strong support for financial institutions to accurately assess enterprise association risks in cross-agency scenarios. The model joint training module fuses the time series model parameters and graph model parameters to determine the global model parameters, realizing the organic combination of time series information and graph structure information in financial data, giving full play to the synergistic advantages of multi-modal data. The global model constructed can comprehensively utilize the cross-agency time series dependence and enterprise association relationship information, more comprehensively and accurately reflect the inherent characteristics of financial data, and improve the performance of the model in financial risk identification and other tasks.

[0070] Further, when the federated aggregation module performs financial time series feature aggregation, it includes:

[0071] The server receives the time interval label attached by each agency client, and maps the financial business cycle to a time encoding vector, where the financial report disclosure quarter corresponds to a high-dimensional activation vector, and the regular cycle corresponds to a benchmark vector.

[0072] The attention calculation unit of the federated time convolution network integrates the time encoding vector and dynamically adjusts the aggregation weight in the following ways: identify the time series feature fluctuation pattern of agencies in the same industry, generate a synergistic enhancement signal for agency groups with a fluctuation phase synchronization degree exceeding the industry threshold; during the financial report disclosure quarter, activate the dedicated feature extraction channel to strengthen the convolution kernel response intensity of the quarterly profit table data; for cross-agency features with dependent relationships, construct a bidirectional attention link to force the association between bank loan data and associated merchant repayment data during aggregation;

[0073] Among them, the time coding vector is to convert the attribute characteristics of the financial business cycle into a numerical vector that can be calculated by the model. Different business cycle characteristics are distinguished by vector dimension and value difference. For example, in the financial scenario, the financial disclosure season (such as the concentration of annual financial report disclosure in March-April) corresponds to a high-dimensional activation vector. By increasing the vector dimension and the numerical value of a specific position, the model can preferentially identify key data such as quarterly income statement and balance sheet; the regular business cycle (such as monthly capital flow) corresponds to the benchmark vector, which maintains stable processing of regular data with basic dimensions and numerical values. This coding method allows the model to automatically "perceive" which business cycle it is currently in, for example, when the system detects a high-dimensional activation vector, it will strengthen the feature extraction of quarterly financial data.

[0074] Synergistic enhancement signal: For the synchronous fluctuation pattern of the time sequence characteristics of the same industry institutions, generate a specific signal to enhance the weight of this kind of institution data in the aggregation process. For example: Assuming that the deposit size of banks generally fluctuates at the end of the quarter, when the system detects that the fluctuation phase synchronization degree of the deposit time sequence characteristics of multiple banks exceeds the industry threshold (such as the fluctuation direction and amplitude tend to be consistent), a synergistic enhancement signal will be generated. This signal prompts the federal time convolution network to pay more attention to the deposit data of this kind of institution when aggregating, for example, when calculating the interbank lending related model parameters, preferentially integrate these synchronous fluctuation deposit data, avoid the aggregation deviation caused by single institution data anomaly.

[0075] Special feature extraction channel: a dedicated data processing path temporarily activated during a specific business cycle (such as the financial disclosure season), used to enhance the feature extraction of key business data. For example: During the financial disclosure season, the system activates the special feature extraction channel, and for the fields such as "net profit" and "operating income" in the quarterly income statement, by adjusting the parameters of the convolution kernel (such as increasing the convolution kernel size and increasing the response weight), the model is more sensitive to the feature extraction of these data. For example, when processing the quarterly income statement data of a certain enterprise, the special channel will enhance the convolution kernel response intensity of the "net profit growth rate" indicator, ensuring that this key indicator occupies a higher weight in the generation of time series model parameters

[0076] In this embodiment, institutions can be classified by different industries, specifically:

[0077] 1. Credit institutions: core features: holding enterprise / personal credit data (such as banks, consumer finance companies). Application of federated learning: joint modeling of credit risk scoring, sharing time sequence characteristics (such as payment records, account flow), capturing default cycle rules through time convolution network.

[0078] 2. Investment institutions: Core characteristics: master capital market data (such as securities companies, fund companies) Federated learning application: Collaborative analysis of enterprise related transactions (such as equity pledge, merger and reorganization), identify related risks in supply chain finance (such as guarantee chain breakage) through graph attention network.

[0079] 3. Risk control institutions: Core characteristics: focus on risk assessment models (such as insurance companies, credit reference agencies). Federated learning application: Share cross-institutional blacklists, multi-head lending data, generate global risk labels through secure multi-party computation, and optimize anti-fraud models.

[0080] 4. Regional small and medium-sized institutions: Core characteristics: small data size but strong local characteristics (such as city commercial banks, rural credit cooperatives). Federated learning application: As a "data provider" to contribute regional characteristic data (such as county enterprise operation data) to exchange global model to improve their risk control ability.

[0081] Based on the current business cycle state, the output layer channel is dynamically selected, and when seasonal risk patterns are detected, long-period convolution kernel group output is enabled to capture cross-year dependent relationship time series model parameters.

[0082] In a specific implementation, the federated time convolution network dynamically aggregates weights based on the following formula:

[0083]

[0084] where, represents the weight coefficient of the cth financial time series feature of the ith financial institution at time point t (such as net profit growth rate in quarterly profit table, asset-liability ratio, etc.) in the federated aggregation process. The weight is used to adjust the importance of the feature in generating the time series model parameters, and the value range is dynamically adjusted according to the business scenario, and the core function is to let the model focus on key features first; is the activation coefficient of the time encoding vector, which is a dynamic parameter reflecting the financial business cycle. In the financial report disclosure season (such as the period when annual financial reports are concentratedly disclosed in March-April every year), its value is determined to be 1.5-2.0 through historical data training, which is used to strengthen the aggregation weight of key data such as quarterly profit table and balance sheet; In the regular business cycle (such as the monthly fund flow period), the baseline value is 1, which maintains stable processing of regular data; is the industry synchronization factor, which is obtained by normalizing the difference between the feature fluctuation phase of institution i and the average phase of the industry through the sigmoid function, with a value range of [0, 1]; This factor is used to measure the synchronization degree of feature fluctuations in the same industry, the smaller the phase difference (i.e. the higher the synchronization degree), the closer to 1, the higher the importance of the institution's features in industry collaborative analysis; The cross-institutional dependency correlation coefficient is used to determine whether there is a stable financial flow dependency relationship between institution i and institution j. When and only when the historical data of the two institutions is stable in the preset interval (such as the time lag of bank loans and business repayments in the supply chain loan scenario is stable in 7-10 days) and holds a valid association certificate verified by secure multi-party computation, is 1, allowing both features to participate in bidirectional attention calculation; otherwise, it is 0, blocking invalid associations. is the actual fluctuation amplitude of the feature of institution i at time t, which is specifically the percentage change of the financial time series data of the institution (such as the quarterly net profit percentage change rate, monthly loan balance percentage change rate, etc.), used to quantify the fluctuation degree of the feature in the time series. For example, if the quarterly end deposit size of a bank increases by 20%, then is 20%; is the industry fluctuation amplitude threshold, which is preset to 15% (unit: %) by statistical analysis of historical data of the same industry. When the actual fluctuation amplitude of the institution feature exceeds the threshold, it indicates that the feature has abnormal or critical fluctuations, and its weight needs to be increased in the aggregation process.

[0085] Based on the above formula, the accurate aggregation of cross-institutional financial time series data is achieved through a three-dimensional dynamic weighting mechanism, solving the problems of missing business cycle features, insufficient industry synergy, and misjudgment of dependency relationships in traditional methods: the time coding activation coefficient automatically identifies key periods such as financial report disclosure season (e.g. every March-April), and increases the aggregation weight of quarterly profit table data by 50%-100% (from the baseline value of 1 to 1.5-2.0). For example, when processing the Q1 financial report of a certain enterprise, the formula will strengthen the convolution kernel response intensity of features such as "net profit growth rate" and "operating income percentage change", avoiding the weakening of key financial data in the regular cycle, and ensuring that the time series model accurately captures the impact of quarterly business fluctuations on risk assessment.

[0086] The industry synchronization degree factor normalizes the phase difference and prioritizes the aggregation of data from the same industry with synchronized fluctuations. For example, when multiple banks experience a synchronized decrease in deposit size at the end of the quarter (with a phase difference less than the industry standard deviation), the industry synchronization degree factor will be close to 1, causing the deposit data weight of these banks to increase, enhancing the model's ability to identify industry-wide risks (such as end-of-quarter liquidity stress), and avoiding interference from single-institution abnormal data in global time series dependency analysis.

[0087] The cross-institutional dependent correlation coefficient is combined with time delay verification (e.g., the loan-repayment time delay of a bank and a supermarket group is stable for 7-10 days) and correlation certificates, and only the real dependent institutions are allowed to participate in the bidirectional attention calculation. For example, the loan-repayment data of a certain commercial bank and a supermarket group are in compliance with the time delay and hold valid certificates, the cross-institutional dependent correlation coefficient = 1, and the fund flow data of the two are forced to be associated when aggregated, so that the model can accurately capture the time-dependent characteristics of "loan amount-repayment ability", avoid mistakenly including the occasional single transaction (without time delay stability) into the calculation, and improve the reliability of credit risk assessment.

[0088] Based on the embodiments provided in the present application, when the federal aggregation module performs financial time series feature aggregation, the financial business cycle is mapped into a time encoding vector (the disclosure quarter corresponds to a high-dimensional activation vector, and the regular cycle corresponds to a benchmark vector), so that the system can accurately identify the feature importance difference of different business cycles. After the attention calculation unit of the federal time convolution network integrates the time encoding vector, it can dynamically adjust the weight for the time series feature fluctuation pattern of the same industry institution: generate a synergistic enhancement signal for the institution group with a fluctuation phase synchronization degree exceeding the industry threshold, ensure the effective synergistic aggregation of the same industry data, activate the dedicated feature extraction channel in the disclosure quarter, strengthen the convolution kernel response intensity of the quarterly profit table data, realize the targeted processing of the key business cycle data, and construct a bidirectional attention link for the cross-institutional features with a dependent relationship. Through the setting of the feature correlation certificate and the certificate verification unit, only the features of the certificate-holding institution group are allowed to participate in the bidirectional attention calculation, ensuring the forced association of dependent relationship data in the aggregation process and avoiding invalid feature interference. Dynamically select the output layer channel based on the current business cycle state, enable the long-period convolution kernel group output when detecting seasonal risk patterns, effectively capture the time series model parameters of the cross-year dependent relationship, improve the extraction ability of the model for the time-dependent features of long-period financial data, and make the aggregation process more in line with the periodic changes and actual needs of the financial business.

[0089] Further, as shown in Figure 2 The cross-institutional graph reasoning in the federal inference module includes:

[0090] S201, the server extracts the guarantee ring topology structure and the supply chain hierarchical relationship from the minimum spanning tree representation, and generates a risk conduction path matrix;

[0091] S202, when the federal graph attention network performs neighbor aggregation, the following operations are performed: a lower limit constraint is applied to the attention coefficient of the nodes in the guarantee ring, so that the mutual attention degree is not lower than a preset business safety threshold; a hierarchical attenuation factor is calculated based on the risk conduction path matrix, and a weighted penalty that is positively correlated with the number of upstream and downstream connections is applied to the risk output of the core enterprise node; a cross-institutional virtual edge is constructed for the nodes with multiple associations to simulate the implicit guarantee relationship in the attention layer.

[0092] In this embodiment, the attention coefficient lower limit constraint sets a minimum threshold for the mutual attention of the nodes in the guarantee ring in the graph attention calculation, ensuring that the risk nodes are not ignored by the model. Example: In the guarantee ring scenario, enterprises X, Y, and Z provide mutual guarantees, and the system sets a lower limit constraint on the attention coefficients of these three nodes (such as not less than 0.3). When calculating the risk of enterprise X, the model must allocate at least 30% of the attention to the associated features of Y and Z (such as guarantee amount, credit rating), avoiding the neglect of the guarantee relationship due to insufficient attention allocation, and accurately capturing the mutual transmission effect of the risk in the guarantee ring, for example, the default risk of Y enterprise will not be underestimated due to insufficient model attention.

[0093] The hierarchical attenuation factor is a parameter that dynamically adjusts the risk transmission intensity according to the supply chain hierarchical relationship and transaction frequency, reflecting the risk attenuation law between nodes at different levels. Example: In the supply chain, core enterprise M directly supplies to downstream enterprise N (one-level transmission), and N supplies to secondary enterprise P (two-level transmission). The basic attenuation coefficient (such as 0.8) is set for the direct upstream and downstream nodes (M-N), and if the transaction frequency between M and N exceeds the industry average, the transmission strengthening mode is started (the attenuation coefficient is increased to 0.9); the cross-level transmission path (M-P) adopts an exponential attenuation strategy, and the attenuation strength is increased by 50% for each additional level (the two-level transmission attenuation coefficient is 0.8 x 0.5 = 0.4). This setting makes the model calculate the default risk of M, and the risk transmission to N is stronger, and the transmission to P is weaker, which conforms to the actual supply chain rule that the risk decreases with the increase of the level.

[0094] The cross-institutional virtual edge is a virtual connection edge constructed between enterprise nodes with multiple indirect associations, used to simulate implicit guarantee relationships that are not explicitly disclosed in reality. Example: Enterprises D and E have no direct guarantee relationship, but they are both important customers of enterprise F, and have formed an implicit guarantee association through F (such as D and E jointly guarantee a loan of F). The system constructs a cross-institutional virtual edge for D and E in the graph attention layer, so that the model can "perceive" this implicit association when calculating. When F has operational risk, the virtual edge will transmit the indirect risk transmission signal between D and E, avoiding the missed judgment of associated risks due to the lack of direct edges, and improving the risk identification ability of complex enterprise relationship networks.

[0095] S203, inject an adversarial test sample to simulate a guarantee chain breakage scenario to verify the risk transmission early warning ability, and dynamically adjust the graph regularization strength according to the verification result.

[0096] Based on the embodiments provided in the present application, when the federal inference module performs cross-institutional graph reasoning, the guarantee ring topology and the supply chain hierarchical relationship are extracted from the minimum spanning tree representation, and a risk transmission path matrix is generated, providing a structured path basis for risk transmission analysis. When the federal graph attention network performs neighbor aggregation, a lower limit constraint is imposed on the attention coefficient of the nodes in the guarantee ring, ensuring that the mutual attention between nodes does not fall below a preset business safety threshold, effectively avoiding the problem of insufficient attention to risk nodes within the guarantee ring; based on the risk transmission path matrix, a hierarchical attenuation factor is calculated, and according to the transaction frequency and default transmission historical data in the supply chain hierarchical relationship, a weighted penalty that is positively correlated with the number of upstream and downstream connections of the core enterprise node is imposed on the risk output of the core enterprise node, while an exponential attenuation strategy is adopted for cross-hierarchical transmission paths, enabling the model to more realistically simulate the transmission rules of risks in different levels and connection relationships; for nodes with multiple associations, cross-institutional virtual edges are constructed to simulate implicit guarantee relationships in the attention layer, making up for the shortcomings of traditional methods in handling implicit association relationships. Injecting adversarial test samples to simulate guarantee chain breaking scenarios and verifying risk transmission early warning capabilities, dynamically adjusting the graph regularization strength according to the verification results, enhances the model's adaptability and early warning capabilities for sudden risk scenarios, and improves the accuracy and robustness of cross-institutional enterprise associated risk identification.

[0097] Bidirectional attention link: Establish a bidirectional association calculation path between cross-institutional features that have a financial flow dependence relationship, and force the association of feature data with business dependence. For example, there is a "loan-repayment" financial flow dependence between a bank and an associated merchant. The server analyzes historical data and finds that the loan data of a certain bank and the repayment data of a certain merchant have a stable time lag of 7-10 days (preset interval), then mark this pair of institutions as a strong dependence group and generate an "inter-institutional feature association certificate". In the aggregation layer of the federal time convolution network, only the bank loan data and merchant repayment data that hold this certificate can enter the bidirectional attention link to achieve forced association calculation. For example, when calculating the loan overdue risk model parameters, the system will preferentially associate the loan time and repayment time series of this pair of institutions, avoiding interference from unrelated institution features.

[0098] Further, the construction of the bidirectional attention link includes:

[0099] The server analyzes the time lag correlation of the aligned bank loan data and merchant repayment data;

[0100] Mark the institution pair with a stable time lag in the preset interval as a strong dependence group, and generate an inter-institutional feature association certificate;

[0101] The inter-institutional feature correlation certificate is an electronic voucher for marking the stable fund circulation dependency relationship between institutions, serving as an access voucher for participating in bidirectional attention calculation. For example, a commercial bank has long-term cooperation with a supermarket chain group, and the supermarket repays the supply chain loan to the bank on the 15th of each month, and the loan and repayment data of the two have a stable time lag within a fixed interval. After the system verifies this stable dependency relationship, a special inter-institutional feature correlation certificate is generated. In the subsequent federated aggregation process, only the bank loan data and supermarket repayment data carrying the certificate can be associated in the bidirectional attention link to ensure that the model accurately captures the time series dependency features of the two, such as the association between loan amount and repayment ability.

[0102] A certificate verification unit is deployed in the aggregation layer of the federated time convolution network to allow only the features of the certificate-holding institution group to participate in bidirectional attention calculation.

[0103] In an optional implementation, the access to the bidirectional attention link is determined based on the following formula:

[0104]

[0105] wherein, is the access identifier for institution i and institution j participating in bidirectional attention calculation, which is a dimensionless Boolean value. When and only when the loan-repayment time lag of the two institutions is within the preset reasonable interval and an effective correlation certificate is held, is 1, allowing its features to enter the bidirectional attention link; otherwise, it is 0, prohibiting participation in calculation; is the actual time lag of the loan data of institution i and the repayment data of institution j, which is calculated by a timestamp alignment algorithm, and the unit is day, reflecting the time interval from loan to first repayment. For example, a bank lends money to a merchant on January 1, 2025, and the merchant repays the loan for the first time on January 8, 2025, which is 7 days; is an indicator function, outputting 1 when the condition in the parentheses is true, and 0 otherwise; 、 is the time lag interval boundary preset for specific business scenarios, with the unit being day. Taking the supply chain loan scenario as an example, combined with the average repayment period in the industry, the is set to 5 days (the shortest reasonable time lag), is set to 15 days (the longest reasonable time lag) to filter institution pairs with time lags that meet business logic; is the inter-institutional correlation certificate status, which verifies whether institutions i and j have a long-term stable business cooperation relationship (such as a continuous 6-month time lag within the preset interval) through secure multi-party computation. If the verification is passed, is 1 (valid certificate), otherwise it is 0 (invalid certificate), which is a dimensionless Boolean value, ensuring that only truly dependent institution pairs participate in the calculation.

[0106] Based on the above formula, the "time lag abnormal interference" and "invalid association mixed" problems in cross-institutional data association are solved through a double condition screening mechanism, ensuring that the federal time convolution network only processes features that are truly business dependent: the indicator function combines a preset time lag interval (such as 5-15 days in the supply chain loan scenario), effectively excluding institutions with time lag abnormalities. For example, if a small and micro enterprise applies for a short-term loan from a bank and the repayment time lag exceeds 30 days (beyond the preset interval),

[0107] The indicator function outputs 0, blocking its data from entering the bidirectional attention link, avoiding misjudgment of the capital flow relationship caused by cross-quarter repayment, and ensuring that the model focuses on effective data that meet the industry repayment period.

[0108] Inter-institutional association certificate Based on historical transaction stability (such as continuous 6-month time lag compliance), it is ensured that the institution pair participating in the calculation has a long-term cooperative relationship. For example, a bank and a local restaurant vendor obtain a valid certificate due to continuous annual on-time repayment (time lag stability within 7-10 days), =1, the loan data and flow data of both can be calculated with bidirectional attention, allowing the model to accurately learn the "loan period - operating cash flow" association pattern; one-time transaction institution pairs (without a certificate) will be automatically intercepted, reducing noise data interference on the model and improving the accuracy of small and micro enterprise credit assessment.

[0109] In the embodiment, the certificate verification unit is deployed in the control module of the federated time convolution network aggregation layer, and is used to check whether the institution features have the permission to participate in the bidirectional attention calculation. For example, when the bank client uploads the loan data and the merchant client uploads the repayment data, the certificate verification unit first checks whether the two hold a common inter-institution feature association certificate. If a small merchant does not have a long-term stable loan relationship with the bank (no certificate), its repayment data will be intercepted by the certificate verification unit when aggregated, and cannot participate in the bidirectional attention calculation of the bank loan data, thereby avoiding the interference of invalid features on the model and ensuring that only institution data that actually exist in business dependence can be used for correlation analysis. Based on the embodiments provided in the application, when constructing the bidirectional attention link, the server analyzes the time lag correlation between the bank loan data and the merchant repayment data after historical alignment, marks the institution pairs with stable time lag in the preset interval as a strong dependence group, and generates an inter-institution feature association certificate. At the same time, the certificate verification unit is deployed in the aggregation layer of the federated time convolution network, and only allows the features of the certificate-holding institution group to participate in the bidirectional attention calculation. This technical solution can accurately screen out institution pairs with stable fund flow dependence relationship, avoid the interference of irrelevant institution feature pairs on attention calculation, ensure that bidirectional attention calculation is only performed between institution features with actual business dependence, improve the pertinence and effectiveness of feature aggregation, and enable the model to more accurately capture the time sequence feature correlation between cross-institutions due to fund flow dependence, thereby providing more reliable dependence relationship feature support for subsequent time sequence model parameter generation.

[0110] Further, the calculation of the hierarchical attenuation factor includes:

[0111] Analyzing the transaction frequency and default conduction historical data in the supply chain hierarchical relationship;

[0112] Setting a basic attenuation coefficient for the direct upstream and downstream nodes, and starting the conduction strengthening mode when the transaction frequency exceeds the industry average;

[0113] Using an exponential attenuation strategy for cross-hierarchical conduction paths, and increasing the attenuation strength by a preset multiple for each additional hierarchical level.

[0114] In a specific embodiment, the supply chain risk conduction hierarchical attenuation factor is determined based on the following formula:

[0115]

[0116] wherein, is the risk conduction attenuation factor from the upstream node k to the downstream node l in the supply chain, reflecting the conduction strength of the risk under different hierarchical levels and transaction frequencies, and the value range is [0, 1]. The closer the value is to 1, the stronger the risk conduction; the closer the value is to 0, the more significant the risk attenuation; is the basic attenuation coefficient of the direct upstream and downstream nodes, preset as 0.8 by statistical industry historical default data, indicating the intensity of risk transmission from the upstream node to the direct downstream node in the absence of high-frequency transactions (for example, when the core enterprise defaults, the first-tier supplier undertakes 80% of the direct risk transmission); is the transaction frequency reinforcement factor, preset as 0.2, used to amplify the risk transmission intensity between high-frequency transaction nodes; when the actual transaction frequency of nodes k and l exceeds the industry average, the reinforcement mechanism is triggered, and the risk transmission intensity is increased by 20% based on the basic coefficient (for example, from 0.8 to 0.96); is the indicator function, outputting 1 when the condition in the parentheses is true, and 0 otherwise; is the actual transaction frequency of node k and l in the statistical period (such as a quarter), unit: times / quarter, obtained by statistical transaction flow between enterprises. For example, the core automobile enterprise and the first-tier tire supplier transact 12 times a quarter, then is 12 times / quarter; is the average transaction frequency of similar node pairs in the same industry, unit: times / quarter, obtained by industry data statistical platform, used to judge whether the transaction of k and l is high-frequency (when the reinforcement mechanism is triggered); is the risk transmission level number, the direct upstream and downstream nodes (such as core enterprise→first-tier supplier) are d=1, one level across (such as core enterprise→second-tier supplier) are d=2, and so on, which is a dimensionless pure counting parameter, used to quantify the hierarchical distance of risk transmission; is the cross-level attenuation multiple, preset as 0.6, indicating that for every additional transmission level, the risk transmission intensity is attenuated by 40% (i.e. the transmission intensity of the current level is 60% of the previous level).

[0117] Based on the above formula, the problem of "one-size-fits-all" risk transmission intensity in traditional methods is solved through a hierarchical sensitive dynamic attenuation mechanism, realizing differentiated modeling of supply chain risk:

[0118] The transaction frequency reinforcement factor triggers the transmission enhancement mode for high-frequency transaction nodes (such as core automobile enterprise and first-tier supplier transacting 12 times a quarter, exceeding the industry average of 8 times), increasing the risk transmission intensity from the basic coefficient 0.8 to 0.96. For example, when the core automobile enterprise defaults, the formula will strengthen its risk transmission weight to the first-tier supplier, making the model more sensitive to the risk rapid spread effect in high-frequency transaction scenarios, avoiding the risk of insufficient risk transmission due to fixed attenuation coefficient, and improving the risk warning ability of the core link of the supply chain.

[0119] The cross-level attenuation factor (preset 0.6) makes the risk transmission intensity significantly attenuate with the increase of the level (for example, the second-level supplier undertakes 48% of the risk, and the third-level supplier undertakes 28.8%). For example, in the new energy vehicle supply chain, the default risk of the core battery manufacturer is strong to the first-level positive material supplier (direct upstream and downstream, d=1), and attenuates to the second-level lithium mine exploiter (across one level, d=2), which conforms to the actual business rule that the risk decreases with the increase of the supply chain level, avoids the model from mistakenly amplifying the risk transmission intensity of the remote level, and improves the risk identification rationality of the complex supply chain network.

[0120] Based on the embodiments provided in the present application, in the calculation process of the level attenuation factor, the transaction frequency and default transmission historical data in the supply chain level relationship are analyzed, the basic attenuation coefficient is set for the direct upstream and downstream nodes, the transmission intensification mode is started when the transaction frequency exceeds the industry average, the exponential attenuation strategy is adopted for the cross-level transmission path, and the attenuation intensity is increased by a preset multiple for each increase of one level. This way of dynamically adjusting the attenuation factor according to the actual transaction frequency and default transmission history enables the model to fully consider the influence of different levels and transaction frequencies in the supply chain on risk transmission: intensifying the risk transmission effect between the direct upstream and downstream nodes of high-frequency transactions, reasonably attenuating the risk intensity of cross-level transmission, and being more consistent with the risk transmission rule in the actual business of the supply chain. This avoids the inaccuracy of the traditional fixed attenuation strategy in simulating the risk transmission of the complex supply chain, and improves the rationality and accuracy of the model in analyzing the enterprise correlation risk in the supply chain level relationship.

[0121] Further, the generation of the adversarial test sample includes:

[0122] Identifying the core enterprise node of the guarantee ring topology in the minimum spanning tree representation;

[0123] Simulating a sudden default event of the core enterprise node, and diffusing the default state along the risk transmission path matrix layer by layer;

[0124] In the present embodiment, the risk transmission path matrix is a structured risk transmission atlas formed by extracting the guarantee ring topology and the supply chain level relationship based on the minimum spanning tree representation of the enterprise correlation relationship. For example, when analyzing the guarantee relationship of enterprises, the system identifies a guarantee ring composed of A enterprise, B enterprise and C enterprise (mutual guarantee) from the minimum spanning tree, and determines that A enterprise is the upstream supplier of B enterprise (supply chain level relationship). The risk transmission path matrix converts these relationships into a calculable matrix form, records the connection mode of the nodes in the guarantee ring (such as A→B→C→A) and the level transmission order in the supply chain (such as A→B). When A enterprise defaults, the matrix can clearly show that the risk may be transmitted to B and C enterprises along the guarantee ring and the supply chain, providing a structured basis for the model to calculate the risk diffusion range.

[0125] A timestamped virtual default event sequence injection test data set is generated.

[0126] According to the embodiments provided in the present application, in the generation process of the adversarial test sample, the core enterprise node of the guarantee ring topology in the minimum spanning tree representation is first identified, then the core enterprise node is simulated to have a default event and the default state is diffused along the risk transmission path matrix layer by layer to generate a timestamped virtual default event sequence injection test data set. This operation can actively simulate the extreme risk scenarios that may occur in reality (such as the sudden default of a core enterprise) in the model training and inference process, and verify the risk transmission early warning ability of the model when facing such scenarios. By dynamically adjusting the graph regularization strength according to the verification result, the model can better adapt to the inference needs under different risk scenarios, enhance the identification and early warning ability of the model to sudden risks such as guarantee chain breakage, and ensure the reliability and effectiveness of cross-institutional graph reasoning in the face of complex risk situations in actual financial business.

[0127] Further, the control of the periodic adaptive output includes:

[0128] A financial seasonal feature library is established to store historical seasonal risk fluctuation feature patterns;

[0129] In the present embodiment, the financial seasonal feature library is a database for storing seasonal risk fluctuation patterns in historical financial data, which is used to compare whether the current time sequence features conform to the seasonal regularity in real time. For example, the library stores the liquidity risk fluctuation feature patterns at the end of each quarter (March, June, September, and December) in previous years, including the fluctuation amplitude of the interbank borrowing rate, the amount of short-term liabilities of enterprises, and other features. When the system detects that the similarity between the current time sequence features and the "quarter-end liquidity tightness pattern" in the library exceeds the threshold, it automatically switches to the long-period convolution kernel group to focus on capturing the liquidity data dependency across years (such as comparing the liquidity indicators in March this year and March last year), while freezing the regular channel's overreaction to short-term fluctuations to ensure that the model accurately identifies seasonal risks.

[0130] Real-time comparison of the similarity between the current time sequence features and the feature library;

[0131] When the similarity exceeds the matching threshold, automatically switch to the long-period convolution kernel group and freeze the regular output channel.

[0132] Based on the embodiments provided in the present application, the control of the periodic adaptive output stores the historical seasonal risk fluctuation feature mode in the financial seasonal feature library, compares the similarity between the current time sequence feature and the feature library in real time, and automatically switches to the long-period convolution kernel group and freezes the regular output channel when the similarity exceeds the matching threshold. This mechanism enables the system to perceive the seasonal risk features in the financial data in real time. When the seasonal risk pattern is detected, the model's feature extraction method is adjusted in a timely manner to capture the long-period dependence relationship across years using the long-period convolution kernel group, while avoiding the interference of the regular output channel on the extraction of seasonal risk features. This adaptive output control strategy improves the model's processing capability for seasonal financial data features, ensuring that the model can extract time sequence features of the corresponding time scale under different business cycle states, and improving the accuracy of the time sequence model parameter generation and the recognition ability of the seasonal risk.

[0133] Optionally, as shown in Figure 3 The present application provides a privacy compliance verification method for cross-institution financial data federated learning modeling, comprising:

[0134] S301, based on the financial business cycle rule, verifying whether the time interval label attached to the financial time sequence feature uploaded by each institution client covers the complete business cycle core data segment;

[0135] S302, calling the dynamically adjusted federated aggregation frequency parameter to verify whether the real-time aggregation frequency meets the preset frequency range of the current business cycle;

[0136] S303, for the time sequence model parameters aggregated by the federated time convolution network, detecting the sensitive feature leakage risk through differential noise disturbance, and verifying the feature extraction interval compliance based on counterfactual reasoning;

[0137] S304, in the process of generating the minimum spanning tree representation by secure multi-party computation, verifying the sensitive field hiding and association relationship disclosure authority of each local subgraph by zero-knowledge proof;

[0138] S305, checking the matching relationship between the dimension of the graph embedding vector generated by the institution client and the complexity of the local subgraph;

[0139] S306, monitoring the attention weight distribution of the federated graph attention network, locating and blocking illegal association edge reasoning;

[0140] S307, detecting the cross-modal illegal association in the fusion process of the time sequence model parameters and the graph model parameters, and reducing the violation coupling degree through feature orthogonalization.

[0141] The privacy compliance verification method for cross-institution financial data federated learning modeling of the application closely surrounds the financial business cycle and the data privacy protection demand, and realizes fine compliance verification and privacy protection in the cross-institution financial data federated learning process. Based on the financial business cycle rule, it is verified whether the time interval label attached to the financial time series features uploaded by each institution client covers the complete business cycle core data segment, ensuring the integrity of data collection and avoiding the omission of key business cycle data, meeting the requirements of data integrity in data use compliance of financial supervision. The dynamically adjusted federated aggregation frequency parameter is called to verify whether the real-time aggregation frequency meets the preset frequency range of the current business cycle, ensuring that the data aggregation process matches the financial business cycle and avoiding the influence of unreasonable aggregation frequency on data processing effect and compliance, improving the standardization and controllability of the data aggregation process.

[0142] For the time series model parameters aggregated by the federated time convolution network, the sensitive feature leakage risk is detected through differential noise disturbance, and the feature extraction interval compliance is verified based on counterfactual reasoning, which can effectively detect and prevent sensitive feature leakage risk in time series model parameters, while ensuring that the feature extraction interval meets the specified compliance requirements, fully protecting data privacy and compliance during model training using time series data. In the process of generating the minimum spanning tree representation by secure multi-party computation, the sensitive field hiding and association relationship disclosure permission of each institution's local subgraph are verified by zero-knowledge proof, ensuring that the sensitive fields are effectively hidden when sharing local subgraph information, and the disclosure of the association relationship meets the respective permission regulations, solving the problem of sensitive information protection and permission control in cross-institution data sharing, ensuring the compliance and security of the data sharing process. The dimension of the graph embedding vector generated by the verification institution client matches the complexity of the local subgraph, ensuring that the graph embedding vector can accurately reflect the structural characteristics of the local subgraph, avoiding data feature loss or redundancy problems caused by unreasonable dimension, and improving the accuracy and effectiveness of graph data processing. The attention weight distribution of the federated graph attention network is monitored to locate and block illegal association edge reasoning, which can timely discover and prevent the model from learning and utilizing illegal association edges during reasoning, preventing risk misjudgment and data abuse problems caused by illegal association reasoning, and ensuring the compliance and accuracy of the graph reasoning process. Cross-modal illegal association in the fusion process of time series model parameters and graph model parameters is detected, and the feature orthogonalization is used to reduce the violation coupling degree, avoiding illegal association of different modal data in the fusion process, ensuring the compliance of multi-modal data fusion, and improving the legality and reliability of global model parameters, providing comprehensive and effective privacy compliance protection for cross-institution financial data federated learning modeling.

[0143] Further, in the process of verifying the time interval label based on the financial business cycle rule, the dynamic compliance control of the timestamp hash alignment strategy is used to realize closed-loop repair, specifically including:

[0144] Establish a binding rule between financial data types and anti-collision hash algorithms: configure SHA-3 algorithm for credit data and BLAKE3 algorithm for high-frequency trading data;

[0145] When it is detected that the institutional client uses a non-bound algorithm, real-time feature transmission is blocked and an algorithm reset instruction is issued;

[0146] Load the time deviation threshold corresponding to the current business cycle type of the system: allow ±4 hours deviation for daily transaction cycle, and allow ±1 working day deviation for quarterly settlement cycle;

[0147] In this embodiment, the time deviation threshold is a timestamp deviation range set according to different financial business cycles, which is used to judge whether the time consistency of the data of the institutional client is compliant. For example, the daily transaction cycle (such as high-frequency stock trading) allows a timestamp deviation of ±4 hours to ensure that the time alignment tolerance of real-time transaction data is low; the quarterly settlement cycle (such as enterprise quarterly financial data reporting) allows a deviation of ±1 working day, considering that there is a certain time window for processing quarterly data. If the timestamp of the quarterly financial report data of an institution is 2 working days later than the standard settlement day (exceeding the threshold), the system will trigger the data repair process to avoid the loss of core data segments (such as quarterly end asset and liability data) due to time deviation.

[0148] If the timestamp deviation of the institution exceeds the time deviation threshold, activate the data repair channel: intercept the original time series data in the deviation period, and generate a compensation feature vector by calling the local aggregation function of the federated time convolution network;

[0149] The data repair channel is an automatic repair mechanism for generating a compensation feature vector by local aggregation for institutional data with timestamp deviation exceeding the threshold. For example, if the timestamp deviation of the quarterly claim data uploaded by an insurance company exceeds the threshold (e.g., 2 days later than the standard time), the system intercepts the original time series data in the deviation period (2 days later than the standard time), calls the local aggregation function of the federated time convolution network, and generates a compensation feature vector containing missing time period claim frequency, amount distribution, etc. The vector is re-submitted for verification after being attached with the correct time interval label, until it covers the complete quarterly core data segment (such as the 10-day concentrated claim data at the end of the quarter), ensuring that the time series model parameter generation does not miss the key period risk features.

[0150] The compensation feature vector is re-submitted to the verification process with the time interval label, and the cycle is executed until the complete business cycle core data segment is covered and detected.

[0151] Based on the embodiments provided in the present application, in the process of verifying the time interval label based on the financial business cycle rule, the dynamic compliance control of the timestamp hash alignment strategy realizes closed-loop repair: the binding rule of financial data type and anti-collision hash algorithm is established, the feature transmission using non-binding algorithm is blocked in real time, and the algorithm reset instruction is issued, so as to ensure the compliance of the hash algorithm in the data transmission process; the time deviation threshold corresponding to the current business cycle type of the system is loaded, the deviation of the agency timestamp is detected, and when the deviation exceeds the threshold, the data repair channel is activated, the original time sequence data in the deviation period is intercepted, and the local aggregation function of the federal time convolution network is called to generate a compensation feature vector, and the compensation feature vector is resubmitted to the verification process until it passes the detection. This closed-loop repair mechanism effectively solves the data compliance problem caused by the timestamp deviation of the agency client and the improper use of the hash algorithm, ensures that the time interval label attached to the financial time sequence features uploaded by each agency can accurately cover the core data segment of the complete business cycle, and guarantees the integrity and compliance of the data in the time dimension, providing a reliable data foundation for subsequent federal aggregation and model training.

[0152] Further, the method further comprises:

[0153] A real-time auditor is implanted in the secure multi-party computation process, and when it is detected that the guarantee relationship edge weight is greater than 40% of the registered capital of the associated enterprise, the over-limit edge is pruned and a violation event snapshot is generated;

[0154] The real-time auditor is a monitoring module implanted in the secure multi-party computation process, which detects in real time whether the guarantee relationship edge weight meets the compliance threshold, and blocks the violation of association leakage. Example: when the guarantee amount of enterprise Q to enterprise R in the local subgraph uploaded by an agency accounts for 45% of the registered capital of R (exceeding the compliance threshold of 40%), the real-time auditor immediately prunes the over-limit edge and generates a violation event snapshot. At the same time, the system deletes the local subgraph of the agency, and if the global risk score change rate exceeds 10% (indicating that the over-limit edge has a significant impact on risk calculation), it is determined that there is a violation of association leakage, the transmission channel of the graph embedding vector of the agency is frozen, and sensitive guarantee information is prevented from being illegally shared, ensuring the compliance of the minimum spanning tree representation.

[0155] The local subgraph of the agency is deleted, and if the global enterprise association risk score change rate is greater than 10%, it is determined that there is a violation of association leakage; the transmission channel of the graph embedding vector of the agency is frozen, and the agency is marked as a high-risk subgraph source;

[0156] Reconstruct the minimum spanning tree representation based on the compliance edge weight threshold: keep the legal association edges after auditing; inject a virtual balancing node into the topological partition broken by pruning to maintain the connectivity of the enterprise risk transmission path.

[0157] The virtual balance node is a virtual node injected in the area where the graph structure is broken due to the cutting of the illegal edge, and is used to maintain the connectivity of the enterprise risk transmission path. For example, after cutting a cross-limit guarantee edge, the subgraph where enterprises S and T are located is disconnected from the global graph structure, forming an isolated partition. The system injects a virtual balance node V to establish a virtual connection edge between V and S and T (the edge weight is 0 and is only used for connectivity), so that the risk transmission path is reconnected to the global network through V. For example, when S defaults, the risk can still be transmitted to other associated enterprises through V, avoiding the loss of risk transmission path due to structure breakage, and ensuring the integrity and accuracy of cross-institutional graph reasoning.

[0158] Based on the embodiments provided in the present application, in the process of constructing the minimum spanning tree, by implanting a real-time auditor in the secure multi-party computation process, the cross-limit edge with a guarantee relationship edge weight greater than 40% of the registered capital of the associated enterprise is cut and a violation event snapshot is generated, and after deleting the local subgraph of the institution, it is determined whether there is a violation of associated leakage according to the global enterprise associated risk score change rate, if there is, the transmission channel of the institution graph embedding vector is frozen and marked as a high-risk subgraph source. At the same time, based on the compliance edge weight threshold, the minimum spanning tree representation is reconstructed, the legal association edge after auditing is retained, and a virtual balance node is injected into the topological partition broken due to cutting to maintain the connectivity of the enterprise risk transmission path. Under the premise of ensuring data privacy compliance, this series of operations effectively identifies and processes illegal association edges, prevents the leakage of sensitive association information, and maintains the integrity of the graph structure and the connectivity of the risk transmission path through the injection of the virtual balance node, ensuring that the minimum spanning tree representation of the global enterprise association can meet the compliance requirements and accurately reflect the risk transmission relationship between enterprises, and improving the compliance and risk identification effectiveness of the model in the cross-institutional graph reasoning process.

[0159] The above is only a preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A cross-institutional financial data federated learning modeling system, characterized in that, include: The federated aggregation module is used by various institutional clients to extract the financial time-series features of financial data, attach time interval labels, and send them to the server. The server aligns the financial time series features of each institution through a timestamp hash alignment strategy, and dynamically adjusts the frequency of federated aggregation based on the financial business cycle. It then uses a federated temporal convolutional network to aggregate the aligned financial time series features and generate time series model parameters that capture global time series dependencies across institutions. The federated inference module is used by the server to generate a minimum spanning tree representation of the global enterprise relationship based on a local subgraph of enterprise relationships constructed by each agency client and a secure multi-party computation strategy. Each institution's client generates graph embedding vectors based on local subgraphs and sends them to the server; the server uses a federated graph attention network to aggregate the graph embedding vectors and combine them with the minimum spanning tree representation to perform cross-institution graph reasoning and construct graph model parameters to identify enterprise association risks; The model joint training module is used by the server to fuse the time series model parameters and the graph model parameters to determine the global model parameters; Cross-agency graph reasoning in the federated reasoning module includes: The server extracts the guarantee loop topology and supply chain hierarchy from the minimum spanning tree representation to generate a risk transmission path matrix. When performing neighbor aggregation, the federated graph attention network performs the following operations: applies a lower bound constraint on the attention coefficient of nodes within the guarantee ring to ensure that their mutual attention is not lower than a preset business security threshold; calculates a hierarchical attenuation factor based on the risk transmission path matrix and applies a weighted penalty to the risk output of core enterprise nodes that is positively correlated with the number of their upstream and downstream connections; and constructs cross-organizational virtual edges for nodes with multiple associations to simulate implicit guarantee relationships at the attention layer. Inject adversarial test samples to simulate a guarantee chain break scenario to verify the risk transmission and early warning capability, and dynamically adjust the graph regularization strength based on the verification results.

2. The cross-institutional financial data federated learning modeling system according to claim 1, characterized in that, When the federated aggregation module performs financial time-series feature aggregation, it includes: The server receives time interval labels attached by each institutional client and maps the financial business cycle into a time encoding vector, where the financial report disclosure season corresponds to a high-dimensional activation vector and the regular cycle corresponds to a baseline vector. The attention computation unit of the federated temporal convolutional network integrates the temporal encoding vector and dynamically adjusts the aggregation weights in the following ways: identifying the temporal characteristic fluctuation patterns of institutions in the same industry, generating collaborative enhancement signals for groups of institutions whose fluctuation phase synchronization exceeds the industry threshold; activating a dedicated feature extraction channel during the financial report disclosure season to strengthen the convolutional kernel response strength of the quarterly profit statement data; and constructing a bidirectional attention link for cross-institutional features with dependencies on fund flows, so that bank loan data and related merchant repayment data are forcibly associated during the aggregation process. The output layer channel is dynamically selected based on the current business cycle status. When a seasonal risk pattern is detected, the output of the long-period convolutional kernel group is enabled to capture the time series model parameters that capture cross-year dependencies.

3. The cross-institutional financial data federated learning modeling system according to claim 2, characterized in that, The construction of the bidirectional attention link includes: The server analyzes the time lag correlation between historically aligned bank loan data and merchant repayment data. For institution pairs whose time delay is stable within a preset range, they are marked as strong dependency groups, and inter-institution feature association certificates are generated. A certificate verification unit is deployed in the aggregation layer of the federated temporal convolutional network, allowing only features from the certificate holder group to participate in bidirectional attention computation.

4. The cross-institutional financial data federated learning modeling system according to claim 1, characterized in that, The calculation of the hierarchical attenuation factor includes: Analyze the transaction frequency and default transmission historical data in the aforementioned supply chain hierarchy; A basic attenuation coefficient is set for direct upstream and downstream nodes, and the transmission enhancement mode is activated when the transaction frequency exceeds the industry average. An exponential decay strategy is adopted for cross-level conduction paths, with the decay intensity increasing by a preset multiple for each additional level.

5. The cross-institutional financial data federated learning modeling system according to claim 1, characterized in that, The generation of the adversarial test samples includes: Identify the core enterprise node in the guarantee ring topology of the minimum spanning tree representation; Simulate a sudden default event at the core enterprise node, and let the default state spread layer by layer along the risk transmission path matrix; Generate a test dataset of virtual default event sequences with timestamps.

6. The cross-institutional financial data federated learning modeling system according to claim 2, characterized in that, Control of the periodically adaptive output includes: Establish a financial seasonality feature database to store historical seasonal risk fluctuation patterns. Real-time comparison of the similarity between current time-series features and the feature library; When the similarity exceeds the matching threshold, it automatically switches to a long-period convolutional kernel group and freezes the regular output channel.

7. A privacy compliance verification method for cross-institutional financial data federated learning modeling, the method being based on the cross-institutional financial data federated learning modeling system of claim 1, characterized in that, include: Based on the financial business cycle rules, verify whether the time interval labels attached to the financial time series features uploaded by each institution's client cover the core data segments of the complete business cycle. Call the dynamically adjusted federated aggregation frequency parameters to verify whether the real-time aggregation frequency meets the preset frequency range for the current business cycle; For the time series model parameters aggregated by federated temporal convolutional networks, the risk of sensitive feature leakage is detected by differential noise perturbation, and the compliance of feature extraction intervals is verified based on counterfactual reasoning. During the process of generating the minimum spanning tree representation through secure multi-party computation, zero-knowledge proofs are used to verify the permissions for hiding sensitive fields and disclosing relationships in the local subgraphs of each organization. The matching relationship between the dimension of the graph embedding vector generated by the verification agency client and the complexity of the local subgraph; Monitor the attention weight distribution of the federated graph attention network to locate and block illegal edge reasoning; Illegal cross-modal associations are detected during the fusion of time series model parameters and graph model parameters, and the degree of illegal coupling is reduced by feature orthogonalization.

8. The privacy compliance verification method for cross-institutional financial data federated learning modeling according to claim 7, characterized in that, In the process of verifying time interval labels based on financial business cycle rules, closed-loop repair is achieved through dynamic compliance control using timestamp hash alignment strategies, specifically including: Establish binding rules between financial data types and collision-resistant hash algorithms: configure SHA-3 algorithm for credit data and BLAKE3 algorithm for high-frequency trading data; When it is detected that the institutional client is using an unbound algorithm, the feature transmission is blocked in real time and an algorithm reset command is issued. Loading system time deviation thresholds for the current business cycle type: ±4 hours deviation is allowed for daily transaction cycles, and ±1 working day deviation is allowed for quarterly settlement cycles; If the timestamp deviation of the institution exceeds the time deviation threshold, the data repair channel is activated: the original time series data of the deviation period is extracted, and the local aggregation function of the federated temporal convolutional network is called to generate a compensation feature vector; The compensated feature vector is resubmitted to the verification process with an added time interval label, and the process is repeated until the core data segment coverage detection of the complete business cycle is passed.

9. The privacy compliance verification method for cross-institutional financial data federated learning modeling according to claim 7, characterized in that, The method further includes: during the minimum spanning tree construction process... A real-time auditor is embedded in the secure multi-party computation process. When the weight of the guarantee relationship edge is found to be greater than 40% of the registered capital of the related enterprise, the excess edge is pruned and a snapshot of the violation event is generated. Delete the local subgraph of the organization. If the change rate of the global enterprise association risk score is >10%, it is determined that there is illegal association leakage. Freeze the embedded vector transmission channel of the organization's graph and mark it as a high-risk subgraph source. The minimum spanning tree representation is reconstructed based on the compliance edge weight threshold: legitimate related edges after auditing are retained; virtual balancing nodes are injected into topological partitions that are broken due to pruning to maintain the connectivity of enterprise risk transmission paths.

Images