Real-time Cross-border Payment Fraud Detection Using Temporal Graph Neural Networks: A Deep Learning Approach
The rapid expansion of digital payments across borders has led to increased risks in the financial system, especially in the fraud process. Traditional methods show limitations in capturing the spatial-temporal patterns inherent in crossing borders. This paper presents a novel Temporal Graph Neural Network (TGNN) approach for real-time financial fraud detection. The proposed system includes a combination of physical-spatial features and a dynamic graph system designed to model structural changes. The architecture employs a multi-head attention mechanism optimized for cross-border payment characteristics, enabling efficient capture of temporal dependencies and spatial correlations in transaction networks. The experiments carried out on two large-scale real-world databases show the effectiveness of our method. The model achieved 99.24% accuracy on Dataset-A (2.8 million transactions) and 98.76% on Dataset-B (1.5 million transactions), outperforming existing methods. The framework maintains robust performance under varying transaction volumes while reducing false positive rates by 37% compared to baseline models. Real-world deployment validates the model's effectiveness in detecting sophisticated fraud patterns while maintaining low computational overhead. The plan shows significant improvements in both detection accuracy and efficiency, making it suitable for use in cross-border payments.
