Real-Time Credit Card Fraud Detection Using Deep Learning, Graph Neural Networks, And Auto Encoders

Abstract

The rapid growth of digital banking and online financial transactions has significantly increased the risk of fraudulent activities, posing serious challenges to traditional fraud detection systems. Conventional machine learning–based approaches primarily rely on isolated transaction features and often fail to capture complex relationships among users, accounts, and transactions. To address these limitations, this work proposes an enhanced fraud detection framework for banking systems using deep learning techniques that combine Graph Neural Networks (GNNs) and Autoencoders. In the proposed approach, banking entities such as customers, accounts, devices, and transactions are modeled as nodes in a graph, while their interactions are represented as edges, enabling GNNs to effectively learn hidden relational patterns. Autoencoders are employed for unsupervised anomaly detection by identifying abnormal transaction behaviors through reconstruction errors. By integrating relational learning with anomaly detection, the framework efficiently handles large-scale and dynamic transaction data and improves the identification of sophisticated fraud patterns, including collusive and multi-hop fraud behaviors. Experimental analysis demonstrates that the proposed hybrid model achieves higher accuracy, improved recall, and reduced false positives compared to traditional machine learning and deep learning models. This approach provides a scalable and robust solution for real-time fraud detection in modern banking environments.