Financial risk control anti-fraud method based on ai model features and federated learning aggregation

CN122798521APending Publication Date: 2026-09-22GUANGDONG YUECAI CREDIT INFORMATION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611071447.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-20
Publication Date
2026-09-22

AI Technical Summary

Technical Problem

[0006]针对现有技术的不足,本发明提供了基于AI模型特征与联邦学习聚合的金融风控反欺诈方法,解决了现有金融风控系统在处理跨机构业务数据时存在数据隐私泄露风险、多模态异构数据难以协同利用,以及传统联邦聚合机制通信开销大、模型收敛效率低,且深度学习风控模型预测结果缺乏内部透明度、无法应对模型概念漂移的问题

Benefits of technology

[0036]1.本发明采用联邦学习架构并结合差分隐私加噪机制,本地节点提取数据均值与标准差进行线性变换后,依据全局敏感度和隐私预算生成拉普拉斯随机噪音并叠加至特征数据中,使原始业务数据留存在本地,仅与中央服务器交互模型参数更新量,在保证多源异构数据统计特征有效性的前提下,阻断了原始数据的反推路径,满足跨机构风控协同的数据隐私合规要求。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122798521A_ABST
    Figure CN122798521A_ABST
Patent Text Reader

Abstract

The application relates to the field of artificial intelligence and data processing, and discloses a financial risk control anti-fraud method based on an AI model and federated learning aggregation. The method relies on a distributed network architecture comprising local nodes and a central server: the local nodes perform differential privacy noise addition on original data to generate feature data. In the model training stage, the local nodes freeze the original network weights, fine-tune the low-rank adaptation parameters, and upload the parameter update amount; the central server generates a binary mask matrix according to the lower quartile threshold, filters the low-contribution-degree parameters, and performs weighted aggregation, and cyclically updates the global model until convergence. The local nodes use the converged model to predict the anti-fraud probability to perform business decisions, use the Shapley additive explanation algorithm to obtain the feature attribution contribution degree, generate a risk control report, and chain it to store the evidence. The application reduces the computing power and communication overhead on the premise of ensuring that the business data does not leave the domain, and provides tamper-proof and interpretable audit credentials.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and data processing, specifically to a financial risk control and anti-fraud method based on the aggregation of AI model features and federated learning. Background Technology

[0002] Utilizing artificial intelligence models for risk control and fraud detection is a standard technique in financial operations. In practice, accurately identifying complex fraud relies on multi-dimensional business characteristic data. However, due to data privacy protection and compliance requirements, data between different financial institutions or business departments cannot be directly merged and shared. This data isolation results in a limited number of sample feature dimensions available to a single institution, making it difficult for the trained risk control model to learn global fraud patterns and limiting the accuracy of identifying new or cross-domain risks.

[0003] Introducing a distributed training framework enables joint modeling without leaving the domain of the original data. However, during the conventional global parameter synchronization process, each participating node needs to frequently upload all local model parameters or gradient matrices. As the risk control model structure becomes more complex, the interaction of all parameters will consume a high amount of communication network bandwidth and exacerbate the computing load on local nodes. Since the distribution of sample data accumulated by each node locally is different, the uploaded parameter updates contain a large amount of redundant information that contributes little to the convergence of the global network. Directly performing global aggregation will reduce the overall synchronization efficiency and model convergence speed.

[0004] In the actual deployment of risk control models, the prediction process based on deep neural networks is an invisible black box, and the final output anti-fraud probability lacks transparency regarding the internal derivation logic. Financial business scenarios typically need to meet internal audit and external regulatory requirements. When executing account blocking or transaction interception, if the risk control system cannot provide quantitative explanations for specific risk factors, it will struggle to handle business appeals and compliance reviews. The actual distribution of business data will change as fraudulent methods evolve and over time. Staticly deployed risk control models are prone to conceptual drift after long-term operation. Without a dynamic retraining mechanism based on real business feedback, the system's effectiveness in intercepting new fraud samples will continuously decline.

[0005] Therefore, this invention proposes a financial risk control and anti-fraud method based on the aggregation of AI model features and federated learning to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a financial risk control and anti-fraud method based on AI model features and federated learning aggregation. It solves the problems of data privacy leakage risk, difficulty in collaborative utilization of multimodal heterogeneous data, high communication overhead and low model convergence efficiency of traditional federated aggregation mechanisms, and lack of internal transparency and inability to cope with model concept drift in the prediction results of deep learning risk control models when processing cross-institutional business data.

[0007] To achieve the above objectives, this invention provides the following technical solution: a financial risk control and anti-fraud method based on AI model features and federated learning aggregation, relying on a distributed network architecture comprising multiple local nodes and a central server, including the following steps:

[0008] Each local node acquires multi-source heterogeneous raw data within the data layer and performs preprocessing operations to generate multi-source heterogeneous feature data.

[0009] Each local node receives the initial global model parameters and the subsequently updated global model parameters from the central server, and combines the global model parameters with multi-source heterogeneous feature data to train the model and obtain the local model parameter update amount.

[0010] The central server performs an aggregation operation on the local model parameter updates received from each local node to update the global model parameters, and then distributes the updated global model parameters to each local node, repeating this process until the model training converges.

[0011] Each local node uses the converged global model parameters to extract global feature vectors from the transaction data to be detected and maps them to anti-fraud probabilities, then executes decision-making actions.

[0012] The system obtains the decision-making action and anti-fraud probability, issues business instructions through the application programming interface, receives the real business tags after manual verification, sends the real business tags back to each local node, and triggers the execution of online incremental training.

[0013] Preferably, each local node performs preprocessing operations to generate multi-source heterogeneous feature data, including: extracting the mean of the structured data of the corresponding local node and the standard deviation of the structured data of the corresponding local node, and performing a linear transformation on the structured data in the multi-source heterogeneous original data to obtain standardized structured data;

[0014] Based on the global sensitivity of the feature query function and the differential privacy budget parameter, a Laplace random noise variable is generated. The Laplace random noise variable is then superimposed on the standardized structured data for numerical replacement, thereby achieving local data desensitization while preserving the statistical characteristics of the data.

[0015] Preferably, each local node combines global model parameters with multi-source heterogeneous feature data to train the model and obtain the local model parameter update amount, including: extracting local representation vectors of multi-source heterogeneous feature data from each encoder branch corresponding to the global model parameters, merging the local representation vectors through vector concatenation operation, and mapping the merged local representation vectors to global feature vectors using a fusion fully connected layer, thereby completing the feature alignment and information fusion of structured, semi-structured and unstructured data.

[0016] Preferably, the global model parameters include a fixed original weight matrix and low-rank adaptation parameters composed of a first low-rank decomposition matrix and a second low-rank decomposition matrix.

[0017] Each local node obtains the local model parameter update, including: during the model forward propagation computation, the input hidden state simultaneously passes through the fixed original weight matrix and the low-rank bypass branch corresponding to the low-rank adaptation parameters.

[0018] The local joint loss function is calculated by combining the global feature vector and the real business label. Each local node uses the local joint loss function value to perform iterative updates on the low-rank adaptation parameters, and extracts the difference in parameter changes as the local model parameter update amount.

[0019] Preferably, the central server performs an aggregation operation on the local model parameter updates uploaded by each local node to update the global model parameters, including: calculating dynamic weights based on the number of local training samples and the model validation quality corresponding to the local node.

[0020] To address the received local model parameter update amount, a zero-norm mask sparsity mechanism is introduced to generate a binary mask matrix and perform a weighted aggregation operation.

[0021] Specifically, the absolute value of the individual parameter element in the local model parameter update matrix is ​​compared with the zero norm mask sparsity threshold corresponding to the target local node. The threshold value is the lower quartile of the set of absolute values ​​of elements in the local model parameter update matrix.

[0022] When the absolute value of an element is greater than the threshold, the corresponding mask element in the binary mask matrix is ​​set to 1; otherwise, it is set to 0.

[0023] The parameters with low contribution are filtered out based on the generated binary mask matrix, and the global model parameter update is calculated and updated accordingly.

[0024] Preferably, the dynamic weights are calculated as follows: the product of the first weight allocation coefficient and the proportion of the number of local training samples is added to the product of the second weight allocation coefficient and the proportion of the model accuracy index of the local node on the independent validation set to obtain the dynamic weights corresponding to the target local node, wherein the first weight allocation coefficient and the second weight allocation coefficient are both positive numbers and their sum is 1.

[0025] Preferably, after the model training converges, the central server traverses the combination of participating nodes, calculates the difference in model utility before and after the node is added, and calculates the Shapley contribution value corresponding to the target local node using the Shapley value algorithm.

[0026] Based on the calculated Shapley contribution value, a federated collaborative incentive allocation is performed for each local node.

[0027] Preferably, each local node uses the converged global model parameters to extract global feature vectors from the transaction business data to be detected and maps them to anti-fraud probabilities, and executes decision actions, including: fusing the first low-rank decomposition matrix and the second low-rank decomposition matrix contained in the low-rank adaptation parameters with the fixed original weight matrix to generate a deployment weight matrix;

[0028] The risk blocking threshold is dynamically calculated based on the set maximum tolerable false positive rate upper limit. That is, the probability threshold variable corresponding to maximizing the true positive rate under the condition of satisfying the maximum tolerable false positive rate upper limit is used as the risk blocking threshold.

[0029] The anti-fraud probability is compared with the calculated risk blocking threshold. Based on the comparison results, corresponding business interception actions are executed, or a decision is made to determine that there is a fraud risk and trigger corresponding account freezing or manual review instructions.

[0030] Preferably, after the decision-making action is performed, the Shapley addition interpretation algorithm is used to fill the baseline values ​​of the missing features with a preset background dataset to simulate the feature missing state, calculate the difference in model prediction output before and after the addition of a single feature, perform reverse decomposition of the anti-fraud probability, and obtain the attribution contribution corresponding to the target feature.

[0031] Preferably, the attribution contribution values ​​are sorted in descending order according to their numerical values, and a preset number of features with the highest attribution contribution values ​​are extracted as key risk factors. The interception timestamp, anti-fraud probability value, and key risk factors are then structurally combined to generate an interpretable risk control report.

[0032] Cryptographic hash functions are used to perform data hashing operations on interpretable risk control reports and transaction data to be tested, extract data digests, and store the data digests in consortium blockchain nodes for solidification and evidence preservation.

[0033] Each of the local nodes stores the transaction data to be detected and its corresponding real business tag as incremental business data into the local database.

[0034] Each local node asynchronously monitors the recognition accuracy on recent incremental business data. When the recognition accuracy is lower than the set warning threshold for several consecutive days, it is determined that the current model has experienced concept drift, and a model retraining request is sent to the central server as an instruction to trigger the online incremental training, thus triggering a new round of federated collaborative training process.

[0035] This invention provides a financial risk control and anti-fraud method based on the aggregation of AI model features and federated learning. It has the following beneficial effects:

[0036] 1. This invention adopts a federated learning architecture combined with a differential privacy noise addition mechanism. After the local node extracts the mean and standard deviation of the data and performs a linear transformation, it generates Laplace random noise based on global sensitivity and privacy budget and adds it to the feature data. This keeps the original business data locally and only interacts with the central server to update the model parameters. While ensuring the validity of the statistical features of multi-source heterogeneous data, it blocks the back-inference path of the original data and meets the data privacy compliance requirements of cross-institutional risk control collaboration.

[0037] 2. This invention introduces a low-rank adaptation parameter and a zero-norm mask sparsity mechanism. During local training, the input state passes through a fixed original weight matrix and a low-rank bypass, and only the low-rank decomposition matrix is ​​fine-tuned iteratively. At the same time, during the central server aggregation stage, the update amount of parameters with low contribution is masked to zero by comparing the lower quartile threshold. While maintaining the global model feature extraction capability, this invention reduces the computational burden of local node fine-tuning of the model, filters redundant parameters, and compresses the communication network overhead between federated nodes.

[0038] 3. This invention utilizes the Shapley addition interpretation algorithm to calculate the difference in predicted output before and after the addition of a single feature, obtains the attribution contribution, extracts key risk factors to generate a risk control report, and simultaneously stores the data hash digest into the consortium blockchain node; at the same time, it monitors the recognition accuracy of incremental data based on real labels, decomposes the prediction results of the neural network into specific feature quantitative impacts, provides risk control audit credentials with tamper-proof characteristics, and can promptly detect model concept drift and trigger adaptive retraining, maintaining the predictive effectiveness of the model in long-term business operations. Attached Figure Description

[0039] Figure 1 This is a diagram illustrating the method architecture of the present invention;

[0040] Figure 2 This is a flowchart of the method steps of the present invention;

[0041] Figure 3This is a flowchart of the preprocessing layer data processing of the present invention;

[0042] Figure 4 This is a schematic diagram illustrating the principle of local model feature fusion and fine-tuning in this invention.

[0043] Figure 5 This is a schematic diagram illustrating the principle of federal collaborative aggregation and evaluation of the present invention;

[0044] Figure 6 This is a flowchart illustrating the online reasoning, decision-making, and compliance audit closed-loop process of this invention.

[0045] Figure 7 This is a curve showing the risk blocking threshold calculation under the false positive rate constraint of the present invention;

[0046] Figure 8 This is a comparison chart showing the convergence of accuracy in multiple iterations of the federated model of this invention. Detailed Implementation

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] See attached document Figure 1 This invention provides a financial risk control and anti-fraud method based on AI model features and federated learning aggregation, relying on a distributed network architecture that includes multiple local nodes and a central server.

[0049] This distributed network architecture is logically divided into a data layer, a preprocessing layer, a federated model layer, a decision layer, and an application layer. Local nodes deploy the data layer and preprocessing layer. The federated model layer includes local model modules and a federated coordination module. Local nodes deploy the local model modules, while the central server deploys the federated coordination module. The decision layer and application layer are configured on the business execution terminals.

[0050] See attached document Figure 2 This invention provides a financial risk control and anti-fraud method based on the aggregation of AI model features and federated learning, including the following steps:

[0051] S10, each local node obtains multi-source heterogeneous raw data from the data layer through the preprocessing layer. The preprocessing layer performs data cleaning and standardization on the multi-source heterogeneous raw data, and uses a differential privacy noise addition mechanism to replace the numerical values ​​of the standardized data to generate multi-source heterogeneous feature data.

[0052] S20, each local node uses the local model module to receive the initial global model parameters sent by the central server. The local model module receives multi-source heterogeneous feature data and performs multimodal feature fusion, outputting a global feature vector. The local node performs low-rank adaptation parameter fine-tuning on the local model module by calculating the local joint loss function and obtains the local model parameter update amount.

[0053] S30, each local node encrypts the local model parameter update and uploads it to the central server. The central server uses the federated collaboration module to perform zero-norm mask sparse aggregation on the received encrypted parameter update to generate the global model parameter update. The central server updates the global model parameters according to the global model parameter update and synchronously distributes the updated global model parameters to each local node. Each local node and the central server repeatedly execute the operations from step S20 to step S30 until the model training converges.

[0054] S40, the local node uses the converged local model module to extract global feature vectors from the real-time input transaction business data to be detected. The decision layer receives the global feature vectors, uses the classification mapping layer of the local model module to map the global feature vectors into anti-fraud probabilities, and matches the corresponding decision actions to the anti-fraud probabilities according to the set risk threshold set.

[0055] S50: The application layer obtains decision actions and anti-fraud probabilities, issues business instructions through the application programming interface, receives real business tags after manual verification, and sends them back to the data layer of each local node, triggering the federated model layer to perform online incremental training.

[0056] See attached document Figure 3 In step S10 above, the local node obtains the multi-source heterogeneous raw data from the data layer through the preprocessing layer and generates multi-source heterogeneous feature data. This step specifically includes the following execution process:

[0057] S101, the local node's data layer classifies the acquired multi-source heterogeneous raw data according to data modal characteristics and writes it to the local database. The multi-source heterogeneous raw data is divided into structured data, semi-structured data, and unstructured data. The data layer stores structured data in the relational database configured on the local node. This structured data includes numerical and categorical transaction amounts, account balances, and credit records. The data layer stores semi-structured data in a non-relational database. This semi-structured data includes transaction logs and user behavior trajectories with time-related characteristics.

[0058] The data layer stores unstructured data in a distributed file storage system. The unstructured data includes text corpora generated from customer service recordings and screenshots of transaction vouchers.

[0059] S102, the preprocessing layer extracts the aforementioned multi-source heterogeneous raw data from the data layer and performs data cleaning and format standardization. For missing and out-of-bounds abnormal values ​​generated during the acquisition of the multi-source heterogeneous raw data, the preprocessing layer performs numerical imputation and filtering based on the overall distribution characteristics of the local dataset. In one specific implementation, the preprocessing layer performs adaptive imputation based on the data field attributes. Specifically, for continuous numerical fields such as transaction amount and account balance, the mean or median of the same type of data in the local node is used for imputation; for discrete categorical fields such as account status, the mode is used for imputation.

[0060] After data cleaning, due to the significant differences in measurement units between different financial fields, directly inputting them into the model can easily cause weight bias. Therefore, the preprocessing layer performs a standardization transformation on the structured data with these measurement differences. The preprocessing layer extracts global statistical parameters from the local dataset and performs a linear transformation on the continuous numerical feature components. The specific calculation formula is as follows:

[0061] ;

[0062] In the above formula, This represents standardized structured data; Represents raw structured data; This represents the mean of the structured data for the corresponding local node; This represents the standard deviation of the structured data corresponding to the local node; This represents a smoothing constant to prevent the denominator from being zero; it is typically set to 10. -5 Minimum value.

[0063] Simultaneously, for semi-structured data, the preprocessing layer extracts business timestamps from the records, aligns them, and reconstructs them into a fixed-interval time-series data format. For unstructured data, the preprocessing layer performs uniform scaling and cropping of image files at the pixel dimension and uniform character encoding format conversion on text files, completing the alignment operation of the data input format.

[0064] S103, the preprocessing layer desensitizes the standardized business-sensitive fields and uses a differential privacy noise-adding mechanism to generate multi-source heterogeneous feature data for model training. For desensitizing fields that clearly identify user identity information, such as ID card numbers and bank card numbers, those skilled in the art can use cryptographic hash functions for one-way encrypted mapping, which will not be elaborated here.

[0065] In distributed collaborative scenarios, to reduce the risk of local data features being inferred backwards in subsequent modeling, the preprocessing layer introduces differential privacy operations based on the Laplace mechanism for standardized continuous numerical features.

[0066] The Laplace mechanism works by injecting random noise following a specific distribution into the true numerical data, making it impossible for external observers to accurately separate a single user's actual draft data from the noise-added result. The preprocessing layer generates random noise variables based on a set privacy protection level and then superimposes these random noise variables onto the standardized structured data for numerical replacement. The specific calculation formula is as follows:

[0067] ;

[0068] In the above formula, This represents the structured feature data after adding noise; This represents standardized structured data; Represents a Laplace random noise variable; symbol This indicates that the probability follows a certain probability distribution; This represents the probability density function of the Laplace distribution; Indicates the global sensitivity of the feature query function; This represents the differential privacy budget parameter.

[0069] Furthermore, global sensitivity The value depends on the maximum possible variation of the feature field to be anonymized in the actual business scenario. Privacy Budget Parameter It is a real number greater than zero, which is preset by the local node according to the privacy protection level expected in the risk control scenario. The smaller the value of this parameter, the greater the variance of the injected noise, and the higher the corresponding level of privacy protection.

[0070] Through the aforementioned format reconstruction and numerical replacement operations, the preprocessing layer can reduce the format differences and privacy leakage risks among multi-source heterogeneous raw data, outputting multi-source heterogeneous feature data that maintains consistency in data distribution and possesses security attributes. This data is then passed to the local model module configured on the local node as the underlying input elements for model forward propagation.

[0071] See attached document Figure 4 In step S20 above, each local node uses its local model module to receive the initial global model parameters sent by the central server and performs multimodal feature fusion and parameter fine-tuning. This step specifically includes the following execution process:

[0072] In step S201, the central server distributes the pre-trained initial global model parameters to each participating local node. The local nodes receive these initial global model parameters and use them to initialize and configure their local model modules. These local model modules are built upon a deep neural network architecture that incorporates multi-layered self-attention mechanisms.

[0073] In one implementation, the local model module can use a pre-trained large model based on the Transformer architecture as a base, which contains a network hierarchy consisting of alternating stacks of multi-head self-attention layers and feedforward neural network layers. To adapt to the computational resource limitations of the local node and reduce subsequent communication overhead, after initialization, the local model module locks the global model parameters of the fully connected layers and attention layers in the original backbone network, so that they remain fixed during subsequent local training and do not participate in gradient updates.

[0074] S202, the local model module receives multi-source heterogeneous feature data from the preprocessing layer input and performs feature extraction using internal parallel encoder branches. The local model module configures a multilayer perceptron encoder for noisy structured feature data, a recurrent neural network encoder for semi-structured data, and a text attention encoder or convolutional neural network encoder for unstructured data.

[0075] For the specific network layer structure and conventional parameter settings of the above-mentioned encoders in different modal data feature extraction, those skilled in the art can make conventional designs in combination with specific data dimensions. These are well-known technologies in the field and will not be elaborated here.

[0076] After each encoder outputs a local representation vector with reduced dimensionality and uniformity (e.g., mapped to a dense vector with the same internal dimension), the local model module merges the local representation vectors through a concatenation operation and maps them to a unified global feature vector using an internal fusion fully connected layer. The specific calculation formula is as follows:

[0077] ;

[0078] In the above formula, This represents the global feature vector generated by the i-th sample in the local node; This represents the weight matrix of the fused fully connected layer; This represents a vector concatenation operation; , and These represent the encoder mapping functions corresponding to structured data, semi-structured data, and unstructured data, respectively. This represents the structured feature data after adding noise; This represents semi-structured data; Represents unstructured data; This represents the bias term for fusing fully connected layers.

[0079] S203, the local model module introduces a low-rank adaptation technique to perform parameter fine-tuning. Specifically, the local model module adds a bypass branch consisting of the product of two low-rank matrices to the side of the fixed linear layers of the backbone network. In principle, to ensure that the output of the bypass branch is zero in the initial stage of fine-tuning and thus does not interfere with the output of the original backbone network, the second low-rank decomposition matrix... Typically, a random Gaussian distribution is used for initialization, and the first low-rank decomposition matrix... Initialization is performed using an all-zero matrix.

[0080] During the forward propagation computation of the model, the input hidden state simultaneously passes through a fixed original weight matrix and a low-rank bypass branch. The outputs of both are added together and then passed to the next network layer. The computational logic is as follows:

[0081] ;

[0082] In the above formula, This indicates the output hidden state of the linear layer; Indicates that the input is hidden; This represents the fixed original weight matrix; This represents the first low-rank decomposition matrix after training convergence. This represents the second low-rank decomposition matrix after training convergence; and the rank of the low-rank decomposition matrix is ​​much smaller than the feature dimension of the original weight matrix. During the fine-tuning phase, the local model module only uses the features of the original weight matrix. and The set is set as trainable local low-rank adaptation parameters, which reduces the number of parameters that need to be updated and optimized for local nodes.

[0083] S204, the local model module outputs the corresponding business fraud prediction probability based on the generated global feature vector, and calculates the local joint loss function by combining it with the real business labels (e.g., using a binary label structure, where 0 represents normal transactions and 1 represents fraudulent transactions). To mitigate data heterogeneity and model bias caused by different financial institutions' different business focuses, this local joint loss function, in addition to the conventional cross-entropy classification loss, superimposes the maximum mean difference distribution alignment loss. Its specific calculation formula is as follows:

[0084] ;

[0085] In the above formula, This represents the local joint loss function value corresponding to the local low-rank adaptation parameters; This represents the standard cross-entropy loss term calculated based on the predicted probability and the true label; This represents the balance coefficient used to adjust the loss ratio; its value can be set between 0.1 and 0.5 depending on the heterogeneity of the historical dataset. This represents the maximum mean difference loss term; this term is used to calculate the distance difference between the global feature vector distribution of the current batch of samples and the historical round reference feature distribution issued by the central server.

[0086] The local node employs a gradient-based optimization algorithm (such as an adaptive moment estimator) to perform backpropagation and iterative updates on the local low-rank adaptation parameters using the calculated local joint loss function value. The conventional derivative calculation process for gradient descent and backpropagation described above can be implemented by those skilled in the art based on conventional neural network training methods, and will not be elaborated upon here.

[0087] After the local node completes the number of training iterations set locally, it extracts the difference between the local low-rank adaptation parameters at this time and the parameter changes in the initial state, and uses it as the local model parameter update amount, ready to execute the subsequent upload operation.

[0088] See attached document Figure 5 In step S30 above, each local node encrypts the local model parameter update and uploads it to the central server. The central server performs parameter aggregation and global model parameter distribution, and repeats this process until the model converges. This step specifically includes the following execution process:

[0089] S301, after obtaining the local model parameter update, each local node encrypts the local model parameter update using a security protection protocol. To enable the central server to perform zero-norm mask sparse operations on the underlying parameters based on absolute value determination while ensuring data privacy, this embodiment adopts Trusted Execution Environment (TEE) technology.

[0090] Each local node uses the TEE public key configured on the central server to encrypt the local model parameter updates and establish a secure network channel. At the same time, each local node extracts the model accuracy metric on the independent validation set for this round, the number of local training samples, and the local feature mean vector representing the global feature distribution, and packages them together with the encrypted parameter updates and uploads them to the central server.

[0091] S302, the federated coordination module deployed on the central server receives the encrypted parameter updates and various indicator parameters uploaded by each local node. The federated coordination module decrypts the encrypted parameter updates within a secure memory isolation area inside the Trusted Execution Environment (TEE), and performs zero-norm mask sparse aggregation under a secure state where the plaintext data is isolated from external systems, generating global model parameter updates. During the aggregation calculation process, the federated coordination module calculates dynamic weights based on the number of samples participated in training by each local node and the model validation quality. The specific calculation formula is as follows:

[0092] ;

[0093] In the above formula, This represents the dynamic weight corresponding to the target local node; Indicates the first weight allocation coefficient; This represents the second weight allocation coefficient; where both the first and second weight allocation coefficients are positive numbers and their sum is 1; This indicates the number of local training samples corresponding to the target local node; This represents the summation symbol; Indicates the traversal index variable; This indicates the number of local training samples for the corresponding local node during the traversal process. Indicates the total number of local nodes participating in the collaboration; This represents the model accuracy metric of the target local node on the independent validation set. This represents the model accuracy metric of the corresponding local node on the independent validation set during the traversal process.

[0094] After determining the dynamic weights, the federated coordination module introduces a zero-norm mask sparsity mechanism to handle parameter updates. The module determines the absolute value of each element in the parameter matrix, generates a binary mask matrix to filter redundant parameters, and performs a weighted aggregation operation based on the dynamic weights. The specific calculation formula is as follows:

[0095] ;

[0096] ;

[0097] In the above formula, This represents the value of the single-unit mask element in the corresponding row and column of the binary mask matrix; Indicates the absolute value symbol; This represents the value of the individual parameter element in the corresponding row and column of the local model parameter update matrix; This represents the zero-norm mask sparsity threshold corresponding to the target local node; its value is the lower quartile of the set of absolute values ​​of the elements in the local model parameter update matrix corresponding to the target local node. This indicates the amount of global model parameter updates generated; Represents the local node index variable; Indicates the total number of local nodes participating in the collaboration; This represents the dynamic weight corresponding to the target local node; This represents the local model parameter update matrix corresponding to the target local node; This represents the element-wise multiplication operation of a matrix; This represents the binary mask matrix corresponding to the target local node.

[0098] In step S303, after obtaining the updated global model parameters, the central server adds them to the existing global model parameters, completing the parameter update iteration for this round. Simultaneously, within the Trusted Execution Environment (TEE), the central server performs a weighted average of the local feature mean vectors uploaded by each local node, generating a historical round reference feature distribution to characterize the global spatial state. The central server then synchronously distributes the updated global model parameters and this historical round reference feature distribution to each local node, allowing them to use it as a benchmark for calculating the maximum mean difference (MMD) distribution alignment loss during subsequent training.

[0099] Each local node and the central server cycle through local training, parameter uploading, and global aggregation operations. After each round of aggregation, the central server calculates the difference in validation accuracy of the global model on the public test set. When this difference is less than a set convergence threshold (e.g., 0.001) for several consecutive rounds, the central server determines that the model training has converged and issues a training stop command to each local node.

[0100] S304. After the model training converges, the central server performs a federal ecological contribution assessment based on Shapley contribution values. The central server uses cooperative game theory to traverse the combinations of participating nodes and calculates the difference in model utility before and after a node is added. The specific calculation formula is as follows:

[0101] ;

[0102] In the above formula, This represents the Shapley contribution value corresponding to the target local node; This indicates a subset of the alliance that does not contain the target local node; This indicates that the subset contains symbols; This represents the complete set of local nodes participating in the collaboration; The symbol for set difference; This represents a single-element set containing only the target local node; This indicates the number of local nodes contained within a subset of the alliance. This represents the factorial operator. Indicates the total number of local nodes participating in the collaboration; This represents the performance benchmark score of the consortium subset after the target local node is added; The symbol for the union of sets; This represents the performance benchmark score of the alliance subset that has not joined the target local node.

[0103] For the actual calculation of the performance benchmark score, the federated collaboration module extracts the equivalent parameter update amount accumulated by each local node in the corresponding federation subset during the historical training process, performs global aggregation, generates a locally converged model, and calculates the area under the curve of the anti-fraud identification result of the locally converged model on the public validation set as the performance benchmark score. After the central server calculates the Shapley contribution value corresponding to each local node, it allocates the corresponding model use rights and benefit settlement coefficients to each local node according to the proportion of the value, thereby realizing the federated collaboration incentive allocation for each local node.

[0104] See attached document Figure 6 In step S40 above, each local node deploys the converged global model, performs anti-fraud reasoning and decision-making on real-time business data, and constructs a closed-loop feedback mechanism. This step specifically includes the following execution process:

[0105] S401, each local node receives the final global model parameters from the central server and updates its local model module accordingly. The online inference phase differs from the training phase, typically having strict requirements for response latency. Therefore, the local nodes perform equivalent matrix fusion of the local low-rank adaptation parameters after training convergence with the original fixed backbone network parameters to generate a deployment weight matrix. The specific calculation formula is as follows:

[0106] ;

[0107] In the above formula, This represents the fused deployment weight matrix; This represents the fixed original weight matrix; This represents the first low-rank decomposition matrix after training convergence. This represents the second low-rank decomposition matrix after training convergence.

[0108] The output tensor after matrix multiplication of the first and second low-rank decomposition matrices has the same row and column dimensions as the fixed original weight matrix. Through the above parameter reparameterization operation, the local node folds the feature representation learned by the side branch into the backbone network, which helps control the computational overhead during online inference. The local node loads the fused deployment weight matrix into the internally configured risk decision module for online execution.

[0109] S402, the risk decision module obtains real-time incremental business data (such as the amount of a single transaction, the transaction timestamp, and the fingerprint of the terminal device), calls the preprocessing layer to perform the same data cleaning and standardization transformation operations as in the historical training phase, and generates dimension-aligned real-time feature vectors.

[0110] The risk decision module inputs the real-time feature vector into the updated network architecture for forward propagation calculation and outputs the anti-fraud probability corresponding to the current business interaction. This probability value is a floating-point number between 0 and 1. The closer the value is to 1, the higher the risk of fraudulent behavior.

[0111] For the activation function calculation and fully connected tensor mapping in the forward propagation process, those skilled in the art can use conventional deep learning inference engines for conventional configuration. The underlying inference execution logic is a well-known technology in this field and will not be described in detail here.

[0112] S403, the risk decision module compares the output anti-fraud probability with the set risk blocking threshold and executes corresponding business interception actions based on the comparison result. The risk blocking threshold is dynamically calculated by the local node based on historical business performance and false positive rate constraints. The specific calculation formula is as follows:

[0113] ;

[0114] In the above formula, This represents the dynamically calculated risk blocking threshold; The symbol represents the variable value to be taken in order to maximize the objective function; The probability threshold variable represents the traversal. The true positive rate is an indicator representing the true positive rate under a given probability threshold. The false positive rate is an indicator representing the false positive rate under a given probability threshold. This indicates the maximum tolerable false positive rate limit set by the local node; those skilled in the art can set this value between 0.01 and 0.05 according to the tolerance for false positives in the risk control scenario.

[0115] Furthermore, considering the case where extreme distribution shifts result in an empty solution set, if no solution exists in the entire (0,1) interval... Probability threshold variable ≤β The risk decision-making module will automatically output a preset hard safety net threshold (e.g., 0.95) as... The value is set to avoid interrupting the algorithm's calculation. When the output anti-fraud probability exceeds this risk blocking threshold, the risk decision module determines that the incremental business data has a fraud risk and triggers the corresponding account freezing or manual review instruction.

[0116] S404: Local nodes collect genuine business results that have been manually reviewed or confirmed by customer complaints, using them as the true labels for incremental business data and storing them in the local database. Local nodes asynchronously monitor the recognition accuracy of the risk decision-making module on recent incremental business data in the background. When the recognition accuracy is lower than the set warning threshold (e.g., 0.75 or 0.80) for several consecutive days (e.g., 3 to 5 consecutive days), the local node determines that the current model has experienced concept drift and sends a model retraining request to the central server, triggering a new round of federated collaborative training to achieve dynamic iteration and continuous evolution of the risk control model.

[0117] See attached document Figure 6 After identifying fraud risks and executing business interception, each local node performs feature attribution analysis on high-risk business data, generates interpretable reports, and executes tamper-proof audit records. This step specifically includes the following execution process:

[0118] In S501, after triggering a business interception command, the risk decision-making module transmits the current incremental business data and its corresponding real-time feature vector and anti-fraud probability to the compliance audit module deployed within the local node. In financial risk control scenarios, business decisions need to meet the model interpretability requirements of regulatory agencies.

[0119] The compliance audit module uses the Shapley addition interpretation algorithm based on cooperative game theory to decompose the output anti-fraud probability in reverse and calculate the independent contribution of each input feature to the final fraud prediction result.

[0120] S502, the compliance audit module constructs a feature-masked conditional distribution, traverses various combinations of input features, and calculates the difference in model prediction output before and after adding a single feature to evaluate the marginal utility of the feature. In the actual model operation, since directly calculating the conditional expectation when features are missing is extremely difficult, the compliance audit module typically uses a pre-set background dataset (e.g., the feature mean vector calculated from historical normal transaction data) to fill in the baseline values ​​of the missing features, thereby simulating the feature missing state and performing forward propagation inference of the model. The specific feature attribution calculation formula is as follows:

[0121] ;

[0122] In the above formula, This indicates the attribution contribution of the target feature; This represents a subset of features that does not contain the target features; This indicates that the subset contains symbols; This represents the complete set of input features used in the model's inference. The symbol for set difference; Represents a single-element set containing only the target features; Indicates the number of features contained within a feature subset; This represents the total number of features contained within the entire set of input features; This represents the factorial operator. This represents the expected conditional prediction after incorporating the target feature; The symbol for the union of sets; This represents the expected conditional prediction without the inclusion of target features; both of the above expected conditional predictions are approximated using the baseline value filling method described above.

[0123] S503, after obtaining the attribution contribution values ​​for all input features, the compliance audit module sorts them in descending order of numerical value. The module then extracts several features with the highest attribution contribution values ​​(e.g., extracting the top 3 or top 5 features based on the system's preset interpretation granularity parameters). These features typically correspond to abnormal states in specific business operations, such as high-frequency trading at night or logins from devices located in different locations, and are marked as key risk factors triggering this interception action. The compliance audit module then structurally combines the interception timestamp, anti-fraud probability value, and key risk factors to generate an interpretable risk control report that complies with regulatory standards.

[0124] S504, to prevent internal personnel from tampering with risk control interception records and to mitigate potential audit risks, the local node uses a cryptographic hash function (such as the SHA-256 algorithm) to perform a data hash operation on the generated interpretable risk control report and the original incremental business data to extract a fixed-length data digest. The local node packages this data digest into an on-chain transaction, broadcasts it, and stores it in a pre-established consortium blockchain node for solidification and notarization.

[0125] For the consensus algorithm, block generation mechanism and hash tree data structure in the consortium blockchain network, those skilled in the art can deploy and configure them based on conventional distributed ledger technology. The underlying evidence storage logic is a well-known technology in the field and will not be elaborated here. Through the above processing, the local node realizes the underlying traceability and trusted auditing of the business decision-making process.

[0126] Specific application examples:

[0127] Assume the distributed network architecture contains 3 local nodes ( ) and 1 central server.

[0128] 1. Data preprocessing and desensitization

[0129] Suppose that node 1 obtains the original structured data. Combined with the extracted mean of the structured data corresponding to the local nodes Standard deviation of structured data corresponding to local nodes Smoothing constant to prevent the denominator from being zero Standardized structured data is obtained through linear transformation. .

[0130] Subsequently, the global sensitivity of the feature query function is set. Differential privacy budget parameters Generate a Laplace random noise variable that follows a distribution. .

[0131] The structured feature data after adding noise was calculated. .

[0132] This data is related to semi-structured data. Unstructured data After processing and performing vector concatenation operations, a global feature vector is generated.

[0133] 2. Local fine-tuning and collaborative aggregation

[0134] In the fine-tuning of node 1, let the conventional cross-entropy loss term be calculated based on the predicted probability and the true label. Maximum mean difference loss term Balance coefficient used to adjust the loss ratio .

[0135] Calculate the local joint loss function value corresponding to the local low-rank adaptation parameters. Node 1 extracts the local model parameter update matrix corresponding to the target local node. .

[0136] During the aggregation phase, the number of local training samples corresponding to the three nodes is known. , , (Total 10000), Model accuracy metric on the independent validation set. , , (Total 2.50). Let the first weighting allocation coefficient be... With the second weight allocation coefficient .

[0137] Calculate the dynamic weight corresponding to the target local node of node 1. .

[0138] The central server takes the quartiles of the absolute values ​​of node 1's parameters as the zero-norm mask sparsity threshold for the target local node. For the individual parameter element values ​​in the corresponding row and column of the local model parameter update matrix. If equal to Then the values ​​of the single-unit mask elements in the corresponding rows and columns of the binary mask matrix are... If equal to ,but Finally, the binary mask matrix corresponding to the target local node is combined. Generate global model parameter update volume .

[0139] 3. Real-time reasoning and auditing

[0140] When used in an online application, the node will train the first low-rank decomposition matrix after convergence. The second low-rank decomposition matrix after training convergence. With a fixed original weight matrix Merge to generate a deployment weight matrix And calculate the blocking threshold.

[0141] As attached Figure 7 As shown, the horizontal axis represents the probability threshold variable of the traversal. The vertical axis represents the indicator values ​​(ranging from 0 to 1), containing two decreasing true positive rate indicators under a given probability threshold. Curve and false positive rate index under given probability threshold conditions Curve. When The curve drops to the maximum tolerable false positive rate limit set by the local node. When the horizontal boundary is reached (assuming at this time) Algorithm Locking The maximum range is used to locate the dynamically calculated risk blocking threshold. When the predicted anti-fraud probability of the business exceeds 0.82, an interception is triggered, and the entire set of input features participating in the model inference is traversed. Combine and output the attribution contribution of the target features. Complete the audit log.

[0142] Experimental verification and effect comparison:

[0143] Validation was performed using 1 million real risk control data points divided into 5 non-independent, identically distributed nodes. The control group used standard federated average full parameter updates (excluding...). ); The experimental group adopted the present invention Optimize and combine Perform mask sparse aggregation.

[0144] Verification results:

[0145] As attached Figure 8 As shown, the horizontal axis represents the number of aggregation rounds, and the vertical axis represents the validation accuracy. The control group (dashed line) was affected by data heterogeneity, fluctuating wildly in the first 60 rounds before stabilizing at 0.83 after the 80th round; the experimental group (solid line) benefited from the maximum mean difference loss term. The distribution alignment and the values ​​of the single-unit mask elements in the corresponding rows and columns of the binary mask matrix. The redundant parameter filtering results in a trajectory that crosses 0.85 in the 40th round and then converges stably to above 0.91 after the 60th round.

[0146] Furthermore, because the experimental group only fine-tuned the first low-rank decomposition matrix after training convergence. The second low-rank decomposition matrix after training convergence The local model parameter update matrix corresponding to the target local node in a single upload The dimensionality is significantly reduced. (Regarding differential privacy budget parameters) Under these conditions, the overall network transmission time of the experimental group was reduced by 85% compared with the control group, which confirms that the scheme reduces communication load while maintaining high model utility.

Claims

1. A financial risk control and anti-fraud method based on AI model features and federated learning aggregation, relying on a distributed network architecture including multiple local nodes and a central server, characterized in that... Including the following steps: S10, each local node acquires multi-source heterogeneous raw data in the data layer and performs preprocessing operations to generate multi-source heterogeneous feature data; S20, each of the local nodes receives the initial global model parameters and the subsequently updated global model parameters issued by the central server, and performs model training by combining the global model parameters with the multi-source heterogeneous feature data to obtain the local model parameter update amount. S30, the central server performs an aggregation operation on the local model parameter update amounts uploaded by each of the local nodes to update the global model parameters, and then sends the updated global model parameters to each of the local nodes, repeating the process until the model training converges. S40, each local node uses the converged global model parameters to extract global feature vectors from the transaction data to be detected and maps them to anti-fraud probabilities, and then executes decision-making actions. S50: Obtain the decision action and the anti-fraud probability, issue business instructions through the application programming interface, receive the real business tags after manual verification, send the real business tags back to each of the local nodes, and trigger the execution of online incremental training.

2. The financial risk control and anti-fraud method based on AI model features and federated learning aggregation according to claim 1, characterized in that, Each of the local nodes performs preprocessing operations to generate multi-source heterogeneous feature data, including: Extract the mean and standard deviation of the structured data of the corresponding local node, and perform a linear transformation on the structured data in the multi-source heterogeneous original data to obtain standardized structured data. Based on the global sensitivity of the feature query function and the differential privacy budget parameter, a Laplace random noise variable is generated, and the Laplace random noise variable is superimposed on the standardized structured data for numerical replacement.

3. The financial risk control and anti-fraud method based on AI model features and federated learning aggregation according to claim 1, characterized in that, Each of the local nodes combines the global model parameters with the multi-source heterogeneous feature data to train the model and obtain the local model parameter update, including: The local representation vectors of the multi-source heterogeneous feature data are extracted by each encoder branch corresponding to the global model parameters. The local representation vectors are merged by vector concatenation operation. The merged local representation vectors are then mapped to global feature vectors by a fusion fully connected layer.

4. The financial risk control and anti-fraud method based on AI model features and federated learning aggregation according to claim 3, characterized in that, The global model parameters include a fixed original weight matrix and low-rank adaptation parameters composed of a first low-rank decomposition matrix and a second low-rank decomposition matrix. Each local node obtains the local model parameter update amount, including: During the model forward propagation computation, the input hidden state is simultaneously passed through a fixed original weight matrix and a low-rank bypass branch corresponding to the low-rank adaptation parameters. The local joint loss function is calculated by combining the global feature vector and the real business label. Each local node uses the value of the local joint loss function to perform iterative updates on the low-rank adaptation parameters, and extracts the difference in parameter changes as the update amount of the local model parameters.

5. The financial risk control and anti-fraud method based on AI model features and federated learning aggregation according to claim 1, characterized in that, The central server performs an aggregation operation on the local model parameter updates received from each of the local nodes to update the global model parameters, including: Dynamic weights are calculated based on the number of local training samples corresponding to the local node and the model validation quality. To address the received local model parameter updates, a zero-norm mask sparsity mechanism is introduced to generate a binary mask matrix and perform a weighted aggregation operation, as shown in the following formula: ; ; In the formula, This represents the value of the single-unit mask element in the corresponding row and column of the binary mask matrix. Represents the absolute value symbol. This represents the value of the individual parameter element in the corresponding row and column of the local model parameter update matrix. The zero-norm mask sparsity threshold for the target local node is represented by the lower quartile of the set of absolute values ​​of the elements in the local model parameter update matrix corresponding to the target local node. This indicates the amount of global model parameter updates generated. Represents the local node index variable. This represents the total number of local nodes participating in the collaboration. This represents the dynamic weight corresponding to the target local node. This represents the local model parameter update matrix corresponding to the target local node. This represents the element-wise multiplication operation of matrices. This represents the binary mask matrix corresponding to the target local node; The global model parameters are updated based on the generated global model parameter update amount.

6. The financial risk control and anti-fraud method based on AI model features and federated learning aggregation according to claim 5, characterized in that, The formula for calculating the dynamic weight is as follows: ; In the formula, This represents the dynamic weight corresponding to the target local node. This represents the first weight allocation coefficient. This represents the second weight allocation coefficient, where both the first and second weight allocation coefficients are positive numbers and their sum is 1; This indicates the number of local training samples corresponding to the target local node. The symbol represents the summation. This indicates iterating over the index variable. This indicates the number of local training samples for the corresponding local node during the traversal. This represents the total number of local nodes participating in the collaboration. This represents the model accuracy metric of the target local node on the independent validation set. This represents the model accuracy metric of the corresponding local node on the independent validation set during the traversal process.

7. The financial risk control and anti-fraud method based on AI model features and federated learning aggregation according to claim 1, characterized in that, After the model training converges, the central server traverses the participating node combinations, calculates the difference in model utility before and after the node is added, and calculates the Shapley contribution value corresponding to the target local node using the Shapley value algorithm. Based on the calculated Shapley contribution value, a federated collaborative incentive allocation is performed on each of the local nodes.

8. The financial risk control and anti-fraud method based on AI model features and federated learning aggregation according to claim 4, characterized in that, Each local node uses the converged global model parameters to extract a global feature vector from the transaction data to be detected and maps it to an anti-fraud probability, then executes decision-making actions, including: The first low-rank decomposition matrix and the second low-rank decomposition matrix included in the low-rank adaptation parameters are fused with the fixed original weight matrix to generate a deployment weight matrix. The risk blocking threshold is dynamically calculated based on the set maximum tolerable false positive rate upper limit. That is, the probability threshold variable corresponding to maximizing the true positive rate under the condition of satisfying the maximum tolerable false positive rate upper limit is used as the risk blocking threshold. The anti-fraud probability is compared with the calculated risk blocking threshold. Based on the comparison result, corresponding business interception actions are executed, or a fraud risk is determined and a corresponding account freeze or manual review instruction is triggered as the decision action.

9. The financial risk control and anti-fraud method based on AI model features and federated learning aggregation according to claim 1, characterized in that, After the decision-making action is executed, the Shapley addition interpretation algorithm is used to fill in the baseline values ​​of the missing features with a preset background dataset to simulate the feature missing state. The difference in model prediction output before and after the addition of a single feature is calculated, and the anti-fraud probability is decomposed in reverse to obtain the attribution contribution corresponding to the target feature.

10. The financial risk control and anti-fraud method based on AI model features and federated learning aggregation according to claim 9, characterized in that, The attribution contribution values ​​are sorted in descending order according to their numerical values. A preset number of features with the highest attribution contribution values ​​are extracted as key risk factors. The interception timestamp, anti-fraud probability value, and key risk factors are then structurally combined to generate an interpretable risk control report. A cryptographic hash function is used to perform a data hash operation on the interpretable risk control report and the transaction data to be detected, and a data digest is extracted and stored in the consortium blockchain node for solidification and evidence preservation. Each of the local nodes stores the transaction data to be detected and its corresponding real business tag as incremental business data into the local database. Each local node asynchronously monitors the recognition accuracy on recent incremental business data. When the recognition accuracy is lower than the set warning threshold for several consecutive days, it is determined that the current model has experienced concept drift, and a model retraining request is sent to the central server as an instruction to trigger the online incremental training, thus triggering a new round of federated collaborative training process.