FL FraDet: Federated Learning for Privacy-PreservingFraud Detection in Mobile Money Systems

Mobile money systems have revolutionized financial inclusion in developing regions,but their rapid growth has introduced significant fraud risks. Traditionalcentralized fraud detection requires aggregating sensitive transaction data, raisingsubstantial privacy concerns and regulatory challenges. Federated Learning(FL) offers a privacy-preserving alternative by enabling collaborative model trainingwithout sharing raw data. However, applying FL to fraud detection faces criticalchallenges, particularly handling Non-IID (non-independently and identically distributed)data across financial institutions with varying transaction patterns andfraud rates. This paper proposes FL FraDet, a new performance-aware aggregationstrategy that weights client contributions based on their fraud detection F1 scoresrather than sample counts. Our key insight is that in imbalanced fraud detectiontasks, model quality should be prioritized over data quantity during aggregation.We evaluate FL FraDet on the PaySim mobile money dataset across multiple experimentalsettings. Results demonstrate that federated learning achieves comparableperformance to centralized baselines (F1=0.9988 vs. 0.9976) under IID conditions.Under Non-IID heterogeneity, FL FraDet outperforms standard FedAvg by 10.1%on average, with improvements of up to 12.9% in moderate label skew scenarios.Importantly, FL FraDet shows particular effectiveness when at least one client possessessufficient fraud samples, enabling performance-aware weighting to amplifyhigh-quality model contributions. Privacy-utility experiments reveal that differentialprivacy with strong guarantees (ϵ ≤ 1.0) maintains perfect fraud detectionperformance, enabling “privacy for free” deployment. Our findings demonstratethat FL FraDet provides an effective, privacy-preserving solution for collaborativefraud detection in heterogeneous mobile money ecosystems.