Federated Learning with Differential Privacy for Fraud Detection: Evaluating Performance Under IID and Non-IID Data Distributions

Abstract

Federated learning allows financial institutions to collaboratively identify fraud without distributing raw transaction data, while differential privacy safeguards individual records from inference attacks. Utilizing a lightweight four-layer neural network that was trained on a 10,000-sample subset of the PaySim mobile money dataset (originally 6.3 million transactions), this paper provides a systematic empirical evaluation of a differentially private federated fraud detection framework. The framework is evaluated in the context of varying data distributions (IID and multiple non-IID schemes), clients' numbers, participation rates, privacy budgets, and data quantity skews. The findings indicate that IID and moderately heterogeneous settings produce high accuracy and F1-scores. Conversely, performance is substantially undermined by severe non-IID partitions, numerous small clients, and extreme data imbalances, particularly when differential privacy noise is implemented. The non-private model's detection capability is largely preserved by intermediate privacy budgets (approximately ε≈1). Notably, membership inference attack success rates remain at or near random guessing (approximately 50% accuracy) in the absence of differential privacy, indicating that the baseline provides a restricted attack surface. In this evaluation setting, the privacy benefit is further confirmed by the fact that differential privacy at ε≈1 further suppresses the attack advantage toward zero. While these results provide practical configuration advice for federated, privacy-preserving fraud detection systems, they should be interpreted within the context of the simplified experimental setup that was implemented.