A Privacy-Preserving Federated Learning Framework for Financial Crime

This paper presents a Federated Learning (FL) framework tailored for decentralized data environments, focusing on preserving data privacy while training models across multiple branches or regions. Synthetic credit card transaction datasets are utilized to mimic real-world scenarios, where sensitive data is distributed across branches. The proposed framework employs encrypted data training at the local level, ensuring privacy through additive encryption techniques. This work introduces a privacy-preserving FL framework using additive encrypted data training technique as additive encryption method. This ensures data privacy while enabling collaborative model training across decentralized environments. The study compares global test accuracy (GTA) of various ML models on a synthetic credit card transaction dataset, demonstrating the superior performance of the Multi-Layer Perceptron (MLP). Additionally, it reveals high local training accuracies (LTA) of 90% to 98% on encrypted data, showcasing the effectiveness of local training. The paper also highlights the importance of model selection for optimal performance and provides practical insights into classifying fraudulent credit card transactions using the proposed FL framework. These contributions enhance the understanding of effective FL implementation while ensuring data privacy and robust model performance.