In recent years, credit card fraud has cost banks and customers a great deal of money. A strong Fraud Detection System (FDS) is therefore necessary to reduce losses for consumers and banks. Our analysis shows that the dataset of credit card transactions is extremely biased, with many fewer examples of fraudulent purchases than of genuine ones. In addition, banks are typically prohibited from sharing their transaction statistics because of concerns about data security and privacy. These issues make it challenging for FDS to identify fraud tendencies and to identify them. In this investigation, we offer a framework in which we label FFD (Federated learning for Fraud Detection) to train a fraud detection model utilizing behavior features with federated learning and convolutional neural networks (CNN) with Greylag Goose Optimization. In contrast to the conventional FDS trained on cloud-centralized data, FFD allows banks to use training data from their local databases to create fraud detection models. Subsequently, a shared fraud detection model is created by combining locally computed updates. Banks can profit collectively from a shared model without exchanging datasets to safeguard the cardholders’ sensitive information. In addition, an oversampling strategy is employed to counterbalance the skewed dataset. We use an extensive set of actual credit card transactions to assess the effectiveness of our credit card FDS system. The findings demonstrate the great accuracy with which each algorithm may be applied to the detection of credit card fraud.
Keywords: CNN, Credit Card Fraud Detection, Efficiency, Federated Learning, Graylag Goose Optimization, Security
