Federated Learning For Fraud Detection: A Review

Abstract

Fraud detection in banking and payments depends on learning from transaction data that institutions are often unable or unwilling to pool, because of privacy, regulatory, competitive, and security constraints. Federated learning (FL) offers an alternative in which institutions train a shared model by exchanging model updates rather than raw records. This review synthesizes selected scholarly works on FL for fraud detection, covering credit-card and payment fraud, financial-crime detection with graph learning, aggregation methods, non-IID data, class imbalance, differential privacy, secure aggregation, poisoning attacks, communication overhead, concept drift, and real-world deployment. Fraud-specific federated studies are compared with foundational FL, privacy, security, and fraud-detection literature. The literature suggests that FL is a plausible route to collaborative fraud detection, but that its benefits in the fraud setting depend on how well it handles extreme class imbalance, heterogeneous institutional data, evolving fraud patterns, and layered privacy and security protections. Many enabling results were obtained in non-fraud settings and should not be assumed to transfer unchanged. Because the review is based on a selected set of publications rather than an exhaustive database search, it is presented as a focused narrative review and not as a systematic review.