Research On Financial Fraud Detection Method Based On Federated Learning

With the rapid development of Mobile Internet technology,new forms of digital financial products such as third-party payment,online lending and consumer loans continue to emerge,which enrich and facilitate people’s daily life,but bring greater fraud risk at the same time.The changeable financial transaction fraud has brought huge losses to financial institutions and consumers.The traditional risk control methods based on statistics and rules can not effectively detect the emerging fraud patterns.Machine learning and deep learning technology provide a new idea for financial fraud detection.The effect of machine learning model usually depends on a large number of comprehensive data sets.However,each financial institution can only observe some cases of fraud.Due to the natural sensitivity and privacy of financial data,the data of different institutions cannot be collected and processed directly.Financial institutions can only utilize limited local data sets for model training.Due to the problem of data island caused by privacy protection and financial supervision,the application of machine learning technology in financial fraud detection scenarios is limited.This paper proposes an algorithm architecture based on Federated Learning(FL),which trains the business sensitive data of various institutions locally,constructs a central anomaly detection model on the premise of ensuring data privacy,breaks the data island,and realizes the fusion learning of multi-source sensitive data models.The main work includes:(1)This paper constructs the architecture of the joint fraud detection system based on the Federated Averaging algorithm and MLP neural network model,carries out detailed experiments,analyzes the performance of the federal model,the local independent training model and data centralized model,which verifies the effectiveness and availability of the Federated Learning method for fraud detection under the condition of privacy protection.(2)Put forward the concept of model contribution of each node,introduce the attention mechanism based on model contribution in the training,and conduct dynamic fusion(Federated model contribution,Fed MC)on the models trained by each node,which greatly improves the precision rate and F1 score of the joint model.The precision rate is improved by nearly 30%,while the F1 score is improved by about 14%.(3)A statistical model of fraud risk characteristics of private data set and federated risk weight aggregation algorithm(Fed RWA)are proposed,which based on the serious imbalance of data categories and the imbalance of the number and severity of fraud samples among multi-institution nodes.In the training,the Synthetic Minority Oversampling Technique(SMOTE)is combined to dynamically balance the private data sets of local nodes,so as to improve the generalization ability of the model.Experiments show that the proposed method can effectively improve the recall rate and G-mean score of the joint model,and can detect more fraud samples with large amounts,so as to bring more application value to financial institutions.