Federated Learning for Credit Card Fraud Detection: A Privacy-Preserving Approach with Smote Optimization

Abstract

Credit card fraud poses significant global challenges, costing financial institutions billions annually while evolving in sophistication. Traditional machine learning approaches for fraud detection face limitations in data privacy and scalability when dealing with distributed transaction data across multiple institutions. This paper presents a novel federated learning framework for credit card fraud detection that addresses these challenges while maintaining detection accuracy. Our approach leverages distributed machine learning across multiple client institutions (ranging from 5 to 260 clients) without requiring direct data sharing, thus preserving privacy. Through extensive experimentation, we demonstrated that our federated model achieved consistent accuracy (99% with 5–50 clients; 95–98% with 100+ clients) while effectively handling class imbalance through SMOTE integration (AUC = 1.00). The system showed particular effectiveness at mid-range client participation (10–50 clients), establishing an optimal balance between detection performance and computational efficiency. Compared to traditional centralized approaches and alternative data balancing methods (undersampling AUC = 0.97, oversampling AUC = 0.99), our federated solution provides superior privacy preservation without compromising fraud detection capability. The results indicate that this framework offers financial institutions a practical, scalable solution for collaborative fraud detection while maintaining strict data confidentiality requirements.