Deep Learning-Based Real-Time Fraud Detection in Digital Payment Systems: A Multi-Feature Behavioural Analysis Framework

Abstract

Digital payment fraud continues to escalate with the rapid expansion of e-commerce and mobile transactions, posing significant financial risks to both consumers and enterprises. This research presents a comprehensive deep learning framework for real-time fraud detection in digital payment systems through multi-dimensional behavioral analysis. The proposed framework integrates temporal transaction patterns, user behavioral sequences, device fingerprinting, and network topology features to construct a robust fraud identification mechanism. Utilizing convolutional neural networks (CNN) for spatial feature extraction and long short-term memory (LSTM) networks for temporal sequence modeling, the system achieves superior detection accuracy while maintaining low false-positive rates. Experimental evaluation on a dataset comprising 2.3 million transactions demonstrates that the proposed framework attains 96.8% detection accuracy with 2.1% false-positive rate, outperforming traditional rule-based systems by 23.4%. The research provides empirical evidence for the effectiveness of multi-feature fusion in enhancing fraud detection capabilities across diverse transaction scenarios. Implementation results indicate substantial improvements in real-time processing efficiency, with average detection latency reduced to 127 milliseconds, making the framework suitable for high-volume production environments.