Privacy-Preserving AI for Financial Fraud Detection: A Federated Learning Approach with Homomorphic Encryption

The exponential growth in digital financial transactions has intensified the need for robust fraud detection systems while maintaining strict privacy compliance. Traditional centralized AI models require raw data aggregation, creating substantial privacy risks and regulatory compliance challenges under GDPR and PCI-DSS frameworks. This paper presents a novel hybrid architecture combining Federated Learning (FL) and Homomorphic Encryption (HE) to enable real-time fraud detection across multiple financial institutions without compromising sensitive customer data. Our system simulates a multi-bank environment where each institution trains local neural network models on rule-based labeled synthetic transaction data. The CKKS homomorphic encryption scheme enables secure aggregation of encrypted model parameters at a central coordination server. Experimental results demonstrate that our privacy-preserving approach maintains detection accuracy while significantly improving fraud recall through strategic class weighting and threshold optimization. The architecture successfully balances privacy preservation, regulatory compliance, and computational performance, offering a scalable solution for collaborative fraud detection in the financial sector.