Secure Federated Learning Frameworks for Risk Detection in Distributed Financial Systems

Abstract

The increasing digitalization of financial ecosystems has amplified both opportunities and risks, particularly in credit assessment and fraud detection. Conventional centralized machine learning models face challenges in data privacy, scalability, and regulatory compliance. This study explores a secure federated learning (FL) framework that integrates privacy-preserving computation using homomorphic encryption (HE) and Bayesian inference mechanisms, specifically adapting the BayesShield approach. The framework also draws on real-time fraud detection strategies merging decentralized intelligence with secure communication. By combining these paradigms, the proposed system addresses challenges of data isolation, adversarial robustness, and decision transparency in distributed financial systems. Results from recent experimental and simulation studies show improved accuracy in credit risk scoring and fraud classification while maintaining compliance with data protection standards such as GDPR. The paper concludes with open challenges and research directions for scalable and secure federated architectures in FinTech applications.