Federated Learning Based Method For Credit Card Fraud Detection

With the development of the mobile internet,fintech based on artificial intelligence and blockchain affects the consumption behavior of each of us and even changes the development mode of the traditional financial industry.At the same time,risks are being passed at a faster speed and in various dimensions.Fully tapping the potential of fintech and building an effective risk control system to improve financial risk control capabilities and efficiency are essential requirements for the development of fintech in the new era.Modern technologies,particularly innovations in machine learning,have been applied to analyze customers' consumption patterns and prevent potentially fraudulent transactions.Statistics show that credit card transaction data is extremely skewed and the number of fraudulent samples is much smaller than the number of normal samples.Moreover,the data used by banks is sensitive because of the highly confidential relationship between financial transaction information and customer data.Due to the unbalanced dataset,it is difficult for fraud detection systems to recognize the patterns of fraud and detect them.On the other hand,due to the data privacy issues,some machine learning models for fraud detection usually only use the internal data collected separately by each bank,which makes it difficult to coordinate large-scale collaborative learning.This thesis proposes a credit card fraud detection method and system with privacy protection based on federated learning which is different from the traditional fraud detection system that trained by dataset in a data center.Federated fraud detection system in this thesis first uses an oversampling method to rebalance the skewed dataset,and then train the local fraud detection model by using the data in the local databases of each bank.Finally,the global shared fraud detection system is constructed by aggregating the parameters(updates)of the underlying fraud detection model of the clients that participate in federated learning.All banks participating in federated learning can benefit from the global sharing model without sharing local datasets.This method not only maintains the security of the data of banks but also protects the privacy of customers of the banks.This thesis evaluated the performance of the federated fraud detection system on a large-scale public credit card transaction dataset,experimental results show that the federated learning based fraud detection system achieves an average of test AUC to 95.5%,which is about 10% higher than traditional fraud detection system.