A Comprehensive Fraud Detection for Credit Card Transactions in Federated Averaging

Credit card fraud costs card issuers billions of dollars each year. Therefore, an effective Fraud Detection System (FDS) is crucial to minimize losses for banks and cardholders. Based on real data collected by the European Credit Card, the dataset is highly skewed, with the number of fraudulent samples significantly lower than legitimate transactions. Due to data security and privacy concerns, different banks are generally not allowed to share their transaction datasets. These challenges not only make it difficult to build traditional credit card fraud detection models but also to detect fraud effectively. In this paper, we evaluate federated learning for fraud detection in identifying illegitimate transactions. Unlike traditional models trained on centralized cloud data, federated learning models allow banks to train fraud detection models using their own local databases. Subsequently, a shared global model is constructed by aggregating locally computed updates of the fraud detection model. This approach enables banks to collectively benefit from a collaborative model without sharing datasets and safeguards cardholders' sensitive information. Experimental results demonstrate that FDS, based on federated learning, achieved an accuracy rate of up to 97% on the European Credit Card (ECC) transactions dataset after over 60 training rounds.