Deep Transformer Architectures for Real-Time E-Payment Fraud Detection Under Class Imbalance (International Journal of Informatics and Communication Technology (IJ-ICT))

Summary

• Global losses from electronic payment fraud are estimated to reach around USD 48 billion globally by 2025.p.1
• Total global e-commerce sales are projected to reach over $7.5 trillion by 2025 and could exceed $9 trillion by 2027.p.1
• The proposed deep transformer architecture achieves an accuracy of 99.82% and an AUC-ROC of 99.56% on the PaySim benchmark dataset.p.1
• The PaySim dataset includes 50,000 transactions and 11 features, where fraudulent transactions account for 3.5% of cases.p.3
• The transformer model architecture comprises approximately 2.8 million parameters, configured with 4 layers and 8 attention heads.p.4
• Model inference requires just over 3.2ms per transaction on a single CPU core, enabling real-time processing of over 300 transactions per second.p.6

Key insights

1. The proposed transformer model achieves an AUC-ROC of 0.984 on the European credit card dataset and 0.974 on the IEEE-CIS dataset.p.5
2. PaySim evaluation results yield 14,865 true negatives, 126 true positives, 7 false positives, and 2 false negatives.p.5
3. GA-GAN uses gated attention to generate synthetic fraud samples for 100 epochs until reaching a 1:5 sample ratio.p.4
4. Focal loss parameters alpha equal to 0.25 and gamma equal to 2.0 were determined as optimal from validation experiments.p.4
5. The European Union AI Act will be fully in force by August 2026, creating regulatory requirements for transparency and human oversight in financial AI decision-making.p.2
6. Feature explanations generated using SHAP achieved a stability index of 0.94 across resampling experiments.p.6