Robust Fraud Transaction Detection: A Two-Player Game Approach

ABA Banking Journal's report 1 , the corporate cost caused by fraudulent activities will exceed $1.495 trillion, which is 1.36$57.147 billion by 2033 [31] . However, traditional fraud detection systems cannot keep up with the fast evolving fraud activities. Specifically, fraudsters are constantly developing new tactics to evade ML-based detection systems [18], [21], leading to more sophisticated and stealthy fraud activities. These tactics are rooted in the deliberately falsified features in fraud transactions, which cause detection systems to misclassify them into benign transactions. Worse still, fraudsters can employ advanced techniques to detect vulnerabilities in the detection system, leading to more effective methods to falsify features. The act of falsifying transaction features to bypass detection mechanisms can be formalized as executing adversarial attacks against targeted detection systems. Yet, there are three distinct characteristics in falsifying the features of fraud activities: (i) the perturbations are unrestricted. Transactions are constituted with monetary features (e.g., Transfer Amount, Register Capital) or temporal features (e.g., Days After Certified), these features are not restricted in small intervals; (ii) the falsification process is resource consuming. Unlike adversarial examples in the image or language field, which only changes pixels or semantics, falsifying features in fraud detection consumes resources (e.g., falsifying the feature of Days After Certified takes substantial time to maintain accounts); (iii) the fraudsters are profit-driven. Since fraudsters aim to get illegal profits from fraud activities as much as possible, the variation of profits exhibits a profound influence on their fraud behaviors (e.g., over 60% of the fraudsters gave up to purchase new accounts