Advancing Fraud Detection Systems Through Online Learning

Tommaso Paladini, Martino Bernasconi de Luca, Michele Carminati, Mario Polino, Francesco Trovò, Stefano Zanero

Abstract: The rapid increase in digital transactions has led to a consequential surge in financial fraud, requiring an automatic way of defending effectively from such a threat. The past few years experienced a rise in the design and use by financial institutions of different machine learning-based fraud detection systems. However, these solutions may suffer severe drawbacks if a malevolent adversary adapts their behavior over time, making the selection of the existing fraud detectors difficult. In this paper, we study the application of online learning techniques to respond effectively to adaptive attackers. More specifically, the proposed approach takes as input a set of classifiers employed for fraud detection tasks and selects, based on the performances experienced in the past, the one to apply to analyze the next transaction. The use of an online learning approach guarantees to keep at a pace the loss due to the adaptive behavior of the attacker over a given learning period. To validate our methodology, we perform an extensive experimental evaluation using real-world banking data augmented with distinct fraudulent campaigns based on real-world attackers’ models. Our results demonstrate that the proposed approach allows prompt updates to detection models as new patterns and behaviors are occurring, leading to a more robust and effective fraud detection system.