Bank intelligent risk control system based on block chain and multi-modal artificial intelligence

Abstract

The invention discloses a bank intelligent risk control system based on a block chain and multi-modal artificial intelligence. A data acquisition module acquires and integrates bank internal data, third-party data and cross-chain data; the edge computing module is provided with a plurality of edge computing nodes, and high-frequency transaction data is processed in real time by using a lightweight model; the block chain network module adopts a relay chain to connect a bank chain and a supply chain, and performs verification based on zk-SNARK zero knowledge proof; the artificial intelligence module constructs a multi-modal risk assessment model, and performs risk assessment in combination with an XGBoost integrated learning algorithm, GNN deep learning and federated learning; and the application decision module performs risk early warning, dynamic processing, visual monitoring and on-chain auditing according to the evaluation result of the artificial intelligence module. According to the scheme, the problems of information isolated island, difficulty in guaranteeing data authenticity, risk identification lag and the like in the existing bank risk control system are effectively solved, and the bank risk prevention and control capability and the market competitiveness are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fintech, and particularly to a bank intelligent risk control system based on blockchain and multimodal artificial intelligence. Background Art

[0002] With the rapid development of fintech, the banking industry is facing increasingly complex risk challenges. The limitations of existing bank risk control systems include: 1. Data fragmentation and poor interoperability: There is a lack of effective coordination mechanisms between different blockchain networks (such as bank consortium chains, supply chain chains, regulatory chains), forming cross-chain islands.

[0003] 2. Insufficient real-time performance: Centralized data processing cannot meet the millisecond-level response requirements of high-frequency trading scenarios.

[0004] 3. Weak privacy protection: There is a risk of sensitive information leakage during the data sharing process, making it difficult to meet the requirements of regulations such as GDPR.

[0005] 4. Low model interpretability: The decision-making process of AI models is opaque, resulting in difficulties in regulatory auditing.

[0006] For example, patent (application number CN 113554505 A) discloses a bank risk control method and device based on blockchain, which relates to the financial field. This method receives sample data provided by each sample data participant through the blockchain, performs data alignment processing, then conducts model training, and generates bank risk control results. However, this patent has insufficient ability to solve cross-chain collaborative verification and multimodal risk assessment.

[0007] And patent (application number CN 113487326 B) discloses a method and device for setting transaction limit parameters based on smart contracts, which relates to the technical fields of blockchain and artificial intelligence. This method includes: obtaining the historical transaction data of a customer in the bank, obtaining multiple attribute values of the customer based on the historical transaction data; determining the risk acceptance value of the customer based on the multiple attribute values of the customer; determining the transaction limit parameters of the customer according to the smart contract of the blockchain, the risk acceptance value of the customer, and the resources of the customer in the bank; and performing corresponding transaction control processing based on the transaction limit parameters. This invention sets transaction limit parameters based on historical transaction data and the smart contract of the blockchain, but this patent does not solve key problems such as multi-chain collaboration, dynamic threshold optimization, and zero-knowledge proof verification. Summary of the Invention

[0008] In response to the above technical problems, the present invention provides a bank intelligent risk control system based on blockchain and multimodal artificial intelligence. The intelligent risk control system combines blockchain technology, multimodal machine learning algorithm, edge computing and federated learning to solve the problems of data islands, lagging risk assessment, insufficient privacy protection and cross-chain collaboration in financial services.

[0009] The present invention is implemented by adopting the following technical solution: a bank intelligent risk control system based on blockchain and multimodal artificial intelligence, including the following modules: Data collection module: collects and integrates internal bank data, third-party data, and cross-chain data; Edge computing module: Set up multiple edge computing nodes and use lightweight models to process high-frequency trading data in real time; Blockchain network module: uses a relay chain based on a cross-chain protocol to connect the bank chain and the supply chain, and verifies based on zk-SNARK zero-knowledge proof; Artificial intelligence module: build a multimodal risk assessment model, combine XGBoost ensemble learning algorithm, GNN deep learning and federated learning to jointly conduct risk assessment; Application decision module: Risk warning, dynamic disposal, visual monitoring and on-chain audit are carried out through the evaluation results of the artificial intelligence module.

[0010] Specifically, the data acquisition module also includes preprocessing the data, specifically including: Data cleaning: fill missing values ​​for time series data and non-time series data respectively, and remove outliers; Feature engineering: Z-score standardization of numerical features; Data dimensionality reduction: PCA principal component analysis is used to reduce feature dimensions and retain the main information. The expression is: = ; Where X is the original data matrix, U and V are orthogonal matrices, and Σ is the singular value matrix.

[0011] Specifically, the missing value filling is as follows: For time series data, the two closest valid data points before and after the missing point are used. and As the boundary, cubic spline interpolation is used to ensure data smoothness. The expression is: , ; Among them, the coefficient , , , Solved by natural spline boundary conditions, Indicates that in a specific interval A cubic polynomial function defined inside for interpolating missing data; Indicates the position of known data points in the time series, Indicates the position points where interpolation is required; For non-time series data, K-nearest neighbor interpolation is adopted, and the formula is: ; where, Is the number of nearest neighbor points, that is, the number of surrounding neighbor points used to calculate the missing value; Is the j-th eigenvalue of the nearest points to the missing value point, Represents the missing data to be filled, which is determined by taking the average of the eigenvalues of its nearest

[0012] points; where, μ is the mean and δ is the standard deviation.

[0012] Specifically, the relay chain based on the cross-chain protocol realizes data interconnection between different blockchain networks through the cross-chain protocol, and combines federated learning to dynamically aggregate multi-chain data features, specifically including: Dynamic weight allocation: Calculate the data freshness, for the data update time interval t of each chain, calculate the decay coefficient, and the calculation formula is: ; where, λ is the decay coefficient, λ = 0.1; Calculate the weight according to the data quality on the chain: ; where, Is the data volume of the i-th chain, Is the data volume of the j-th chain, Is the data freshness of the i-th chain, Is the data freshness of the j-th chain; Multi-core learning fusion: Fusion of heterogeneous data feature kernel matrices: ; ( 0, ) ; where, Is the feature sum matrix of the i-th chain, Optimized by cross-validation.

[0013] Specifically, the construction of the multi-modal risk assessment model includes: XGBoost static feature modeling: Refine the objective function, introduce Focal Loss to solve the problem of class imbalance, and the expression is: ; where, L is the loss function, Is the true label, that is, the actual class to which the sample belongs, is the predicted probability, that is, the probability value that the model sample belongs to the positive class; by and reduce the weight of classification samples to alleviate class imbalance; Split gain optimization. Aiming at the sparsity of financial data, improve the gain calculation. The improved gain expression is: ; where is the original gain; LSTM time series modeling and lightweighting: Based on the edge computing node and lightweight model of the edge computing module, extract time series features, expressed as: ; where the number of model parameters is compressed to 30% of the original version, and the response time < 20ms; represents the hidden state at time step t on the edge computing node, which contains information from previous time steps and is used for the calculation of the current time step, is the input feature at time step t, that is, the data input to the LSTM model at the current time point; Graph Neural Network GNN Association Analysis: Integrate structured transaction data and unstructured social network data to construct a customer-merchant association graph, which is conducive to GNN mining potential gang fraud patterns. The expression is: ; Node feature contains multi-modal data of transactions, social and device fingerprints; Model training: Use cross-entropy to calculate the loss function ; Use the Adam optimizer to update the model parameters: ; η is the learning rate, and are the first-order and second-order momentum estimates of the gradient.

[0014] Specifically, the federated learning includes dynamic federated learning and privacy protection, specifically including: Federated aggregation optimization: Shapley value contribution estimation: Used to quantify the contributions of banks A and B in the joint anti-money laundering model. For N participating parties, enumerate all subsets ; Calculate the marginal contribution as: ; is the AUC of the model for subset S; Adaptive differential privacy: Input the local model gradient , calculate the gradient sensitivity , generate Gaussian noise: , the total number of training rounds T = 10, and each round consumes , the total budget ; Output the secure gradient after adding noise , upload it to the federal server; Based on the risk score distribution of historical data by zk-SNARK zero-knowledge proof, use kernel density estimation (KDE) to calculate the dynamic threshold, expressed as: Verify ; Calculate the privacy protection score: The calculation formula is: , It means that first perform a certain hash transformation H on the original data D, and then input the transformed data into the XGBoost model for calculation to obtain a result related to the risk score. It means that first perform a hash transformation H on the graph data structure, and then input the transformed data into the graph neural network GNN model for calculation to obtain another result related to the risk score.

[0015] Specifically, the multi-modal risk assessment model further includes model optimization and hyperparameter tuning, specifically: Use Bayesian optimization to tune the model hyperparameters: ; Among them, is the objective function (such as model accuracy), and θ is the hyperparameter combination; Use Gaussian process to model the objective function: ; Among them, is the mean function, is the covariance function; Introduce L2 regularization in model training to prevent overfitting: ; Use the early stopping strategy to stop training when the validation set loss no longer decreases.

[0016] Specifically, the joint risk assessment specifically includes: Build a scoring card model based on logistic regression: ; Among them, is the customer default probability, is the model parameter; Convert the probability to a risk score: ; Among them, , A and B are constants; Use the weighted average method to fuse the prediction results of multiple models: ; Among them, is the weight of the i-th model, .

[0017] Specifically, the blockchain network module further includes: Adopt an improved PBFT algorithm to improve the transaction processing speed: ; Introduce the sharding technology to divide the blockchain network into multiple sub - networks and improve the parallel processing ability; Use the Merkle tree to compress and store data: , where H is the hash function, and A, B, C, D are data blocks.

[0018] The beneficial effects of the present invention are as follows: The present invention provides an end - to - end intelligent risk control system, realizing a full - process closed - loop of cross - chain data collaboration - real - time risk assessment - privacy protection - dynamic optimization. Using blockchain technology to build a bank risk control consortium chain to achieve the sharing and collaboration of risk information; using smart contract technology to realize the automatic collection, verification, encryption and on - chain operation of risk data; using artificial intelligence technology to deeply mine and analyze the on - chain risk data, build a risk assessment model, and realize the real - time assessment and early warning of customer credit risk, market risk, operation risk, etc.; using blockchain technology to establish a risk disposal mechanism to improve the efficiency and transparency of risk disposal. Through cross - chain collaboration, edge computing, zero - knowledge proof and multi - modal fusion, realize the trusted sharing of data, real - time risk assessment and privacy protection. The system supports multiple scenarios such as credit card anti - fraud and supply chain finance, improving the detection accuracy by 40% and reducing the manual intervention cost by 90% compared with the traditional solution. The present invention can effectively solve the problems existing in the existing bank risk control systems, such as information silos, difficulty in guaranteeing data authenticity, and lag in risk identification, and improve the bank's risk prevention and control ability and market competitiveness; it has the following effects: (1)Improve data utilization rate: Cross - chain collaboration and federated learning increase the risk assessment coverage rate by 50%.

[0019] (2)Millisecond - level real - time response: Edge computing reduces the transaction processing delay to less than 20ms.

[0020] (3)Privacy and compliance: zk - SNARK and differential privacy meet the GDPR requirements, reducing the data leakage risk by 95%.

[0021] (4)Dynamic adaptability: Reinforcement learning reduces the false alarm rate of the system by 30% in extreme market environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0023] Figure 1 It is the system five - layer structure and data flow architecture diagram in the embodiment of the present invention; Figure 2Flow chart of cross-chain collaborative verification in an embodiment of the present invention; Figure 3 Structural diagram of a multi-modal GNN model in another embodiment of the present invention; Figure 4 Schematic diagram of state federated learning parameter aggregation in an embodiment of the present invention. Detailed implementation manners

[0024] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Components of the embodiments of the present invention generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0025] It should be noted that like reference numerals and letters denote like items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0026] The following combines the attached Figures 1 to 4 , and details some implementation manners of the present invention. Without conflict, the following embodiments and the features in the embodiments may be combined with each other.

[0027] The present invention proposes a bank intelligent risk control system based on blockchain and multi-modal artificial intelligence. In a preferred embodiment, the system is as Figure 1 shown and includes the following modules: Data acquisition module: Acquire internal bank data, third-party data, and cross-chain data, and integrate them; Edge computing module: Set up multiple edge computing nodes to process high-frequency trading data in real time using lightweight models; Blockchain network module: Use a relay chain based on a cross-chain protocol to connect the bank chain and the supply chain, and perform verification based on zk-SNARK zero-knowledge proof; Artificial intelligence module: Build a multi-modal risk assessment model, combine XGBoost ensemble learning algorithm, GNN deep learning, and federated learning to jointly perform risk assessment; Application decision module: Perform risk early warning, dynamic disposal, visual monitoring, and on-chain auditing through the evaluation results of the artificial intelligence module.

[0028] In one embodiment, the system is divided into five layers: Data acquisition layer: Integrate internal bank data, third-party data (such as social behavior, IoT sensors), and cross-chain data.

[0029] Edge computing layer: The lightweight model processes high-frequency trading data in real time.

[0030] Blockchain network layer: The cross-chain protocol (relay chain) connects multiple chains and supports zk-SNARK verification.

[0031] Artificial intelligence analysis layer: Multi-modal models (XGBoost, GNN, federated learning) jointly conduct risk assessment.

[0032] Application layer: Risk warning, dynamic handling, visual monitoring, and on-chain auditing. Data flow: Data collection → Edge computing → Cross-chain and on-chain → AI analysis → Application decision-making.

[0033] The specific technical implementation of each part of the system is as follows: I. Data preprocessing: Data preprocessing is the foundation of the risk control system, ensuring the quality and consistency of the input data.

[0034] (1) Data cleaning: Missing value filling: For time series data (such as transaction records), using the two nearest valid data points before and after the missing point and as the edges, cubic spline interpolation is used to ensure data smoothness: , ; Among them, the coefficients , , , are solved through natural spline boundary conditions.

[0035] For non-time series data (such as customer attributes), K-nearest neighbor (KNN) interpolation is used: .

[0036] Outlier removal: Outliers are removed using the 3δ principle (three times the standard deviation method): If , then remove ; where μ is the mean and δ is the standard deviation.

[0037] (2) Feature engineering: Standardize numerical features (Z-score standardization): .

[0038] (3) Data dimensionality reduction: Use principal component analysis (PCA) to reduce the feature dimension and retain the main information: ; Among them, X is the original data matrix, U and V are orthogonal matrices, and Σ is the singular value matrix.

[0039] II. Cross-chain feature aggregation: Achieve data interoperability between different blockchain networks (bank consortium chain, supply chain chain, regulatory chain) through cross-chain protocols (such as relay chain or atomic swap), and dynamically aggregate multi-chain data features in combination with federated learning. For example, cross-chain integrate the transaction data of supply chain finance and bank credit data to generate a joint risk assessment model.

[0040] (1) Dynamic weight allocation: Calculate the data freshness, for the data update time interval t of each chain, calculate the decay coefficient: (λ is the decay coefficient, λ = 0.1) Calculate the weight according to the data quality on the chain (such as data volume, update frequency): .

[0041] (2) Multi-kernel learning (MKL) fusion: Fusion of heterogeneous data feature kernel matrices: ; ( 0, ) ; Among them, is the feature sum matrix of the i-th chain, Optimized by cross-validation.

[0042] III. Multimodal risk assessment model: The risk assessment model is the core of the system, and an integrated learning algorithm (such as XGBoost) and a deep learning algorithm (such as LSTM, KNN) are combined to evaluate the credit risk of customers.

[0043] (1) XGBoost static feature modeling: Refinement of the objective function: Introduce Focal Loss to solve the problem of class imbalance: ; Among them, L is the loss function (such as logarithmic loss), through and Reduce the weight of classification samples and alleviate class imbalance (such as the proportion of fraud samples < 1%).

[0044] Split gain optimization: Aiming at the sparsity of financial data, improve the gain calculation: .

[0045] (2) LSTM time series modeling and lightweighting: Deploy edge computing nodes at the data acquisition layer to process high-frequency trading data (such as device fingerprints and geographical locations) in real time. Use a lightweight LSTM (Tiny-LSTM) model to extract temporal features and reduce cloud transmission latency.

[0046] Among them, the number of model parameters is compressed to 30% of the original version, and the response time is < 20ms.

[0047] (3)Graph Neural Network (GNN) Association Analysis: Integrate structured trading data and unstructured social network data to construct a customer-merchant association graph, which is conducive to GNN mining potential gang fraud patterns.

[0048] ; Node features Contain multi-modal data of transactions, social interactions, and device fingerprints.

[0049] (4)Model Training: Adopt the cross-entropy loss function: ; Use the Adam optimizer to update the model parameters: ; where η is the learning rate, and are the first and second moment estimates of the gradient.

[0050] IV. Dynamic Federated Learning and Privacy Protection (1)Federated Aggregation Optimization: Shapley Value Contribution Estimation: Used to quantify the contributions of Banks A and B in the joint anti-money laundering model.

[0051] For N participants, enumerate all subsets .

[0052] Marginal Contribution Calculation: ; is the model AUC of subset S.

[0053] (2)Adaptive Differential Privacy (Gradient Protection): Protect the bank's local training parameters (such as gradients) from being reverse-inferred.

[0054] Input Data: Input the local model gradient , calculate the gradient sensitivity: ; Generate Gaussian noise: ; The total number of training rounds T = 10, and each round consumes , the total budget .

[0055] Output result: Secure gradient after adding noise , and upload it to the federal server.

[0056] (3)Zero-knowledge proof verification zk-SNARK verification: Based on the risk score distribution of historical data, use kernel density estimation (KDE) to calculate the dynamic threshold: Verification .

[0057] (4)Privacy-preserving score: .

[0058] V. System optimization (1)Federated learning: On the premise of protecting data privacy, use federated learning technology to achieve multi-party data collaborative modeling.

[0059] ; Among them, is the local model parameter of the i-th participant, is the data volume of the i-th participant.

[0060] (2)Edge computing: Introduce edge computing at the data acquisition layer to improve data preprocessing efficiency.

[0061] .

[0062] (3)Model update: Regularly update the risk assessment model to adapt to the changing risk environment: .

[0063] VI. Model optimization and hyperparameter tuning (1)Hyperparameter tuning: Use Bayesian optimization to tune the model hyperparameters: ; Among them, is the objective function (such as model accuracy), and θ is the hyperparameter combination.

[0064] Use Gaussian Process to model the objective function: ; Among them, is the mean function, is the covariance function.

[0065] (2)Regularization and early stopping: Introduce L2 regularization in model training to prevent overfitting: ; Use the Early Stopping strategy to stop training when the validation set loss no longer decreases.

[0066] VII. Risk Score Calculation (1) Scorecard Model: Build a scorecard model based on logistic regression: ; where is the customer default probability, are the model parameters.

[0067] Convert the probability to a risk score: ; where , A and B are constants.

[0068] (2) Multi-Model Fusion: Use the weighted average method to fuse the prediction results of multiple models: ; where , A and B are constants.

[0069] VIII. System Performance Evaluation (1) Evaluation Metrics: Accuracy: ; Precision and Recall: , ; F1 Score: ; AUC-ROC Curve: Evaluate the classification performance of the model at different thresholds.

[0070] (2) Stress Testing: Simulate the system performance under extreme scenarios (such as economic recession, market volatility): .

[0071] IX. Blockchain Performance Optimization (1) Consensus Algorithm Optimization: Adopt an improved PBFT (Practical Byzantine Fault Tolerance) algorithm to improve the transaction processing speed: .

[0072] Introduce sharding technology to divide the blockchain network into multiple sub-networks and improve the parallel processing ability.

[0073] (2) Storage Optimization: Use a Merkle Tree to compress and store data: ; Among them, H is a hash function, and A, B, C, and D are data blocks.

[0074] Taking credit card anti-fraud as an example, the following operations are performed based on this system: S1: Data collection: Transaction data (time, amount, merchant category), device fingerprint, geographical location.

[0075] S2: Edge computing preprocessing: Tiny-LSTM real-time detects abnormal transaction sequences (such as 5 large-value transactions within 10 minutes).

[0076] S3: Cross-chain verification: Obtain third-party blacklist data through the relay chain, and zk-SNARK verifies its authenticity.

[0077] S4: Risk assessment XGBoost static score (70) + GNN gang fraud score (20) = comprehensive score 90.

[0078] S5: Risk handling: The smart contract automatically freezes the account and notifies the regulatory agency Taking supply chain finance risk control as an example, the following operations are performed based on this system: S1: Cross-chain data integration: Bank consortium chain (credit data) + supply chain (logistics data) → federated learning joint modeling.

[0079] S2: GNN correlation analysis: Construct a supplier-distributor-logistics network to identify false trade cycles.

[0080] S3: Dynamic threshold adjustment: Reinforcement learning automatically adjusts the risk threshold according to market fluctuations (such as reducing the threshold by 10% during an economic crisis).

[0081] For the foregoing embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that this application is not limited by the described action sequence, because according to this application, some steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily required by this application.

[0082] In the above embodiments, the basic principles, main features and advantages of the present invention are described. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, any modifications and changes made by those skilled in the art that do not depart from the spirit and scope of the present invention should fall within the protection scope of the appended claims of the present invention.

Claims

1. A bank intelligent risk control system based on blockchain and multimodal artificial intelligence, characterized in that Includes the following modules: Data collection module: collects and integrates internal bank data, third-party data, and cross-chain data; Edge computing module: Set up multiple edge computing nodes and use lightweight models to process high-frequency trading data in real time; Blockchain network module: uses a relay chain based on a cross-chain protocol to connect the bank chain and the supply chain, and verifies based on zk-SNARK zero-knowledge proof; Artificial intelligence module: build a multimodal risk assessment model, combine XGBoost ensemble learning algorithm, GNN deep learning and federated learning to jointly conduct risk assessment; Application decision module: Risk warning, dynamic disposal, visual monitoring and on-chain audit are carried out through the evaluation results of the artificial intelligence module.

2. The bank intelligent risk control system based on blockchain and multimodal artificial intelligence according to claim 1, wherein, The data acquisition module also includes preprocessing the data, specifically including: Data cleaning: fill missing values ​​for time series data and non-time series data respectively, and remove outliers; Feature engineering: Z-score standardization of numerical features; Data dimensionality reduction: PCA (Principal Component Analysis) is used to reduce the feature dimension and retain the main information. The expression is as follows: ; where X is the original data matrix, U and V are orthogonal matrices, and Σ is the singular value matrix.

3. The bank intelligent risk control system based on blockchain and multimodal artificial intelligence according to claim 2, characterized in that, The missing value filling is specifically as follows: For time series data, using the two nearest valid data points before and after the missing point and as boundaries, cubic spline interpolation is adopted to ensure data smoothness, and the expression is: ; Among them, the coefficient , , , is solved by natural spline boundary conditions, represents a cubic polynomial function defined within a specific interval , and is used for interpolating missing data; represents the positions of known data points in the time series, represents the position points where interpolation is required; For non-time series data, K-nearest neighbor interpolation is used, and the formula is: ; where is the number of nearest neighbor points, that is, the number of surrounding neighbor points used to calculate the missing value; is the feature value of the j-th point among the k nearest points to the missing value point, represents the missing data to be filled, which is determined by taking the average of the feature values of its k nearest points; The rejection of the outliers is specifically as follows: The outliers are rejected according to the principle of the three - standard - deviation method. If , then is rejected; where μ is the mean value and δ is the standard deviation.

4. The bank intelligent risk control system based on blockchain and multimodal artificial intelligence according to claim 1, characterized in that, The relay chain based on the cross-chain protocol realizes data intercommunication between different blockchain networks through the cross-chain protocol, combined with the characteristics of dynamically aggregating multi-chain data in federated learning, specifically including: Dynamic weight allocation: Calculate the data freshness, for the data update time interval t of each chain, calculate the attenuation coefficient, and the calculation formula is: ; where λ is the attenuation coefficient, λ = 0.1; Calculate the weight according to the data quality on the chain: ; where is the data volume of the i-th chain, is the data volume of the j-th chain, is the data freshness of the i-th chain, is the data freshness of the j-th chain; Multi-core learning fusion: Fusion of heterogeneous data feature kernel matrices: ; where Optimized by cross-validation.

5. The intelligent risk control system for banks based on blockchain and multimodal artificial intelligence according to claim 1, wherein, The construction of the multimodal risk assessment model includes: XGBoost Static Feature Modeling: Refine the objective function, introduce Focal Loss to solve the problem of class imbalance, and the expression is: ; where L is the loss function, is the true label, that is, the actual class to which the sample belongs, is the predicted probability, that is, the probability value that the model sample belongs to the positive class; by and reduce the weights of classification samples to alleviate class imbalance; Split gain optimization. For the sparsity of financial data, the gain calculation is improved, and the improved gain expression is: ; where is the original gain; LSTM Time Series Modeling and Lightweighting: An edge computing node and a lightweight model based on an edge computing module, extracting time series features, are expressed as: ; among which, the number of model parameters is compressed to 30% of the original version, and the response time < 20ms; represents the hidden state at time step t on the edge computing node, which contains information from previous time steps and is used for the calculation of the current time step. is the input feature at time step t, that is, the data input to the LSTM model at the current time point; Graph Neural Network (GNN) Association Analysis: Integrate structured transaction data and unstructured social network data to construct a customer-merchant association graph, which is conducive to GNN mining potential gang fraud patterns. The expression is as follows: , node features include multi-modal data of transactions, social interactions, and device fingerprints; Model training: The loss function is calculated using cross-entropy. ; The model parameters are updated using the Adam optimizer: ; η is the learning rate, and are the first and second moment estimates of the gradient.

6. The bank intelligent risk control system based on blockchain and multimodal artificial intelligence according to claim 1, characterized in that The federated learning includes dynamic federated learning and privacy protection, specifically including: Federated Aggregation Optimization: Shapley Value Contribution Estimation: Used to quantify the contributions of Bank A and Bank B in the joint anti-money laundering model. For N participants, enumerate all subsets ; Calculate the marginal contribution as: ; is the model AUC for subset S; Adaptive Differential Privacy: Input Local Model Gradient , calculate the gradient sensitivity , generate Gaussian noise: , the total number of training rounds T = 10, and each round consumes , the total budget ; Output the secure gradient after adding noise , upload it to the federated server; Based on the zk-SNARK zero-knowledge proof of the risk score distribution based on historical data, calculate the dynamic threshold using kernel density estimation (KDE), expressed as: ; Calculate the privacy protection score: The calculation formula is , indicates that first, a certain hash transformation H is performed on the original data D, and then the transformed data is input into the XGBoost model for calculation to obtain a result related to the risk score. indicates that first, the graph data structure is subjected to the hash transformation H, and then the transformed data is input into the graph neural network GNN model for calculation to obtain another result related to the risk score.

7. The bank intelligent risk control system based on blockchain and multimodal artificial intelligence according to claim 1, characterized in that, The multimodal risk assessment model also includes model optimization and hyperparameter tuning, specifically: Use Bayesian optimization to tune the hyperparameters of the model: ; where is the objective function, and θ is the hyperparameter combination; Modeling the objective function using Gaussian processes: ; where is the mean function, is the covariance function; Introduce L2 regularization in model training to prevent overfitting: ; Use the early stopping strategy to stop training when the validation set loss stops decreasing.

8. A bank intelligent risk control system based on blockchain and multimodal artificial intelligence according to claim 1, characterized in that, The joint risk assessment specifically includes: Build a scoring card model based on logistic regression: ; where is the customer default probability, are the model parameters; Convert probability to a risk score: ; wherein, , A and B are constants; The prediction results of multiple models are fused using the weighted average method: ; where is the weight of the i-th model, .

9. The bank intelligent risk control system based on blockchain and multimodal artificial intelligence according to claim 1, characterized in that, The blockchain network module also includes: Adopt an improved PBFT algorithm to improve the transaction processing speed: ; Introducing sharding technology to divide the blockchain network into multiple sub-networks to improve parallel processing capabilities; Using a Merkle tree to compress and store data: , where H is a hash function, and A, B, C, and D are data blocks.

Images