Federated Learning Approaches for Fraud Detection in International Payment Systems
Main Article Content
Abstract
Although fraud detection methods based on machine learning (ML) have gained momentum, many organizations still rely on rule-based detection methods. Additionally, publicly available datasets are not representative of the types of fraud most commonly faced by the organizations. Therefore, industry partners can benefit from a jointly trained ML model capable of detecting patterns shared across banks or across cross-border payments. However, organizations face regulatory constraints when sharing transaction data or account details with industry peers. Federated learning (FL), an approach that enables collaborative ML without data sharing, represents a potential solution to this problem. Within the context of international payments, this work discusses common types of fraud and considers how FL can help alleviate these burdens. A specific federated learning framework aligned with fraud detection workflows is proposed. The approach supports both vertical and horizontal architectures, incorporates techniques for handling heterogeneous data distributions, and promises to minimize sensitive data sharing. A safeguard associated with the peer-to-peer architecture is the avoidance of a single point of failure, whereas a weakness arises from the need to trust communications made between institutions of the same jurisdiction. Dense anomaly signals within a given jurisdiction can improve soundness by assisting region-based model updates.