Deep Learning Approaches for Intelligent Fraud Detection in Digital Banking Transactions

Abstract

The rapid development of digital banking has not only enhanced convenience for users but also amplified the challenge of protecting against fraud, which remains both expensive and harmful for banks, customers, and partner organizations. Fraud detection systems addressing these issues with large-scale and field-wide transaction records are extremely helpful. However, developing effective and practical fraud-detection systems poses three major challenges. First, defining a learning target based on business or field requirements is crucial for achieving a truly useful system. Second, support from business expertise is essential for understanding features that are difficult to extract automatically, particularly those concerning certain fraud types. Finally, high-quality and genuinely large training sets are difficult to obtain.

The employement of Deep Learning as the main tool in fraud detection is particularly suitable for three reasons: (1) the availability of a truly large-scale transaction dataset, (2) wide enough temporal coverage of the dataset to capture both stable and evolving user behavior, and (3) the repetition characteristic of fraud types that permits formal modeling of fraud scenarios. A supervised-learning approach has been used to detect common types of fraudulent transactions at individual point-in-time locations. A recurrent deep-learning model has been designed to identify more complex fraud types based on human-usage characteristics over time. Furthermore, transaction-level features, engineered on the underlying transaction data with the support of domain expertise, have been employed to help in detecting various types of fraud including text-related and risk-related frauds.