Credit card fraud detection based on federated graph learning

With the widespread popularity of credit card payments, cases of credit card fraud have increased, driving the continuous development of fraud detection technology. Utilizing machine learning to train models for identifying fraudulent transactions has become a primary method for preventing credit card fraud. However, due to the high privacy of transaction data, the training process of credit card fraud detection (CCFD) models also faces the problem of data silos where transaction data between financial institutions are isolated from each other. Moreover, traditional machine learning models only focus on the features of transaction data but ignore the connection between transaction data and transaction data, which makes it difficult for existing credit card fraud detection to cope with escalating fraud tactics. To address the above problems, we propose a CCFD model based on federated graph learning by combining federated learning (FL) and graph neural networks (GNN). Specifically, we have designed a framework for collaborative training of CCFD models among multiple financial institutions. This framework facilitates the joint training of CCFD models by coordinating multiple financial institutions while ensuring data security. In addition, we propose a graph construction algorithm based on weighted feature similarity to map the local datasets of financial institutions into a transaction graph representation for model training. This approach delves into mining connections between transaction data to enhance the performance of the CCFD model. In order to counter fraudulent transactions spanning multiple financial institutions, we propose a graph extension algorithm based on convolutional feedforward generative (CFG) networks. Through federated training of a convolutional feedforward neural network (CFNN) and a generative adversarial network (GAN), this algorithm extends the transaction graph to reinforce connections between financial institution data. Moreover, we propose an adaptive model aggregation method based on cosine similarity to enhance the performance after FL model aggregation. Finally, the simulation results demonstrate that our proposed model achieved the highest recall and AUC values across all test datasets compared to the baseline model. Particularly, on the IEEE-CIS dataset, our proposed model’s recall improved by 11.87% to 33.9% over the baseline, and the AUC increased by more than 3% compared to the baseline algorithm.