Towards an Adversarial Model for Fraud Detection Systems
Résumé
Fraud detection systems (FDS) must contend with several intertwined challenges: class imbalance, robustness against adaptive adversaries, black-box constraints, and need for explainability. While each of these challenges has been addressed individually in the literature, few works offer an integrated approach capable of operating under realistic adversarial settings. These solutions often rely on Logistic Regression to predict fraudulent behavior, and on data augmentation techniques in an indirect or ad hoc manner, aiming to approximate real-world conditions without fully capturing their complexity. In this study, we introduce JDA-NC-FDS, a novel fraud detection model that tackles these 4 issues. First, we assess the comparative performance of Logistic Regression and XGBoost within an iterative adversarial learning setup. Second, we propose to rely on the Jacobian Data Augmentation instead of the traditional SMOTE. In addition, we enhance JDA to incorporate categorical features, a critical yet underexplored type of dimension in fraud detection. Experimental results demonstrate that JDA-NC-FDS improves model performance across multiple adversarial training rounds, especially in terms of PR-AUC, while preserving explainability and adaptability.