Deep transformer architectures for real-time e-payment fraud detection under class imbalance

Abstract

E-commerce has become a rapidly expanding channel, leading to more e-commerce payment transactions and an increased threat for advanced fraudulent activities. New frauds include cross-channel attacks, adversarial behavior and synthetic identities, which are challenging to rule-based detection. In this paper, a deep learning framework for detecting e-payment fraud in real-time is proposed. The model integrates self-attention, explainable AI, class balancing using GANs, and a transformer-based architecture to enable transparent decision-making. The solution tackles major issues in fraud detection, such as class imbalance, evolving transaction trends, and the requirement for interpretability in compliance with regulations like the EU AI Act. Experiments on benchmark datasets demonstrate good results, with an accuracy of 99.82%, and an AUC-ROC of 99.56%. Based on the results, the use of transformer-based models with the help of synthetic data generation and explainability can enhance the reliability and transparency of e-payment fraud detection systems.