Deep Learning-Based Transaction Intelligence for Real-Time Financial Fraud Detection and Risk Mitigation

Abstract

With annual losses resulting from cybercrime now reported at more than US$600 billion, financial institutions absolutely cannot afford to ignore transaction-related fraud. The spate of data breaches combined with the growing sophistication of analysis now renders reactive fraud detection wholly inadequate. Mitigation requires a shift toward guarded environments and real-time transaction monitoring systems designed explicitly around surveillance, and regulation mandates that transactions be evaluated for risk at the moment they are executed. Current research on real-time anomaly detection seeks to build a wide range of models capable of detecting even the rarest and most subtle of classes. The methods used rely heavily on synthetic instances, and are often validated on either the closest adjacent class or on a label-to-label basis. Although some have drawn on combinations of machine and self-supervised techniques to good effect, the use of feedback loops has generally been overlooked.

Complete transaction streams have been found to yield more information than individual transactions, but the focus has nevertheless remained on individual temporal points. Recurrent architectures in particular are well suited to the temporal class. High-level understanding of architecture behavior can simplify the transition to a more effective search for patterns across the entire stream. Recurrent architectures such as long short-term memory or gated recurrent unit, enhanced with attention, support description at different levels of temporal abstraction and indeed across the whole stream. Real-time streaming-inference systems must satisfy additional considerations surrounding end-to-end timing, operational processes, and resilience if the full risk-scoring functionality is to be recreated in a form that fits naturally in a transaction-monitoring context.