Defense Mechanism to Thwart Model Poisoning on Non-IID Data based Federated Learning for Credit Fraud Detection System

The ever-growing sophistication of credit fraud, encompassing methods like identity theft, necessitates robust detection mechanisms. While machine learning algorithms analyze transaction data for fraud patterns, they struggle with complex schemes, fairness, and privacy concerns. This paper proposes a federated learning approach to address these challenges, leveraging the inherent heterogeneity of transaction data. By partitioning data based on merchant categories and geographical regions, this approach enables collaborative model training tailored to each category’s unique characteristics without compromising data privacy. Moreover, federated learning adapts models to diverse transaction contexts, effectively addressing the non-IID nature of the data and enhancing fraud detection across different segments. Furthermore, countermeasures against model poisoning attacks are incorporated to ensure the integrity and efficacy of the overall system.