Skip to content
Conference Open access

Adversarially-Robust and Explainable Reinforcement Learning for Fraud Detection in Digital Banking

Sep 2026 · 2026 10th International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-SMAC) · pp. 1961-1967 · 0 citations · 25 references

Abstract

Reinforcement learning (RL) has recently emerged as a viable adaptive replacement for traditional rule-based and supervised methods for detecting fraud perpetrated against banks that operate online. Nevertheless, RL-based fraud detection systems have always been evaluated in static, non-aggressive environments; hence, we cannot tell the true effectiveness of these systems. To remedy the situation, we propose a self-play system based on adversarial RL, which consists of a defender agent trained to identify transactions as either legitimate or fraudulent and an attacking agent trained to change fraudulent transactions in such a way that they will not be identified as fraudulent. To demonstrate the system’s capabilities, we used the dataset from the PaySim mobile-money simulator and demonstrated that even though the defender agent trained in a standard way achieved quite good results in identifying transactions as legitimate or fraudulent (92.8% precision, 89.7% recall, F1 = 0.912), it does not allow to avoid being successfully attacked in 51.3% percent of cases. Collecting adversarial examples over several rounds eliminates this problem and decreases the evasion rate to 22.4%, while the precision on clean data drops to 72.6%. This situation shows that there is a trade-off phenomenon between robustness and accuracy. We also indicate that a traditional gradient boosting classifier produces nearly ideal metrics for clean data, yet is attacked more effectively (71.6%) than both RL methods, revealing the misleading nature of the said clean data accuracy, which is not a good indicator for evaluating adversarial robustness in this case. Finally, we introduce SHAP-based explainability for the last defender to get per-transaction and overall feature attribution and overcome the unaccountability issue of the previous RL fraud detection systems.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.