AI Fraud Detection Framework Integrates U.S. Banking Regulations

A new framework aimed at enhancing AI-based financial fraud detection in U.S. banks has been introduced, embodying principles from various regulatory guidelines. The RGF-AFFD framework employs a three-tier governance structure, streamlining model development, validation, and oversight. Researchers tested six architectures using two substantial datasets, with the LSTM+XGBoost ensemble achieving an impressive ROC-AUC score of 0.9289 and an F1 score of 0.6360, demonstrating a notable 6-to-1 benefit-cost ratio. Among the models, XGBoost exhibited superior stability over time through evaluations involving ablation, drift, SHAP interpretability, and insights into fairness principles according to BISG standards.