Federated Reinforcement Learning for Adaptive Fraud Behavior Analytics in Digital Banking

Abstract

The rapid growth of digital banking has been paralleled by increasingly sophisticated fraud attempts that adapt to detection mechanisms. Traditional centralized fraud detection models often face challenges such as data privacy concerns, scalability limitations, and delayed adaptability to emerging fraud patterns. To address these issues, this study proposes a Federated Reinforcement Learning (FRL) framework for adaptive fraud behavior analytics in digital banking. The framework enables multiple financial institutions to collaboratively train fraud detection agents without sharing sensitive customer data, thereby preserving privacy and regulatory compliance. By leveraging reinforcement learning, the model continuously adapts to dynamic fraud strategies through feedback-driven policy optimization. Experimental results demonstrate that the proposed FRL approach achieves superior detection accuracy, reduced false positives, and faster adaptation to novel fraud patterns compared to conventional machine learning and federated learning baselines. This research highlights the potential of FRL as a scalable and privacy-preserving solution for combating financial fraud in the era of decentralized and intelligent banking ecosystems.