Federated Learning for Privacy-Preserving Mobile Money Fraud Detection

Abstract

A case study deploying federated learning across mobile money agents to detect fraudulent transaction patterns without centralising sensitive data.

Keywords

Full Text

Mobile money fraud detection typically requires pooling transaction data from multiple agents into a central model, raising privacy and regulatory concerns in jurisdictions with strict data-localisation rules. This case study deployed a federated learning pipeline across 40 mobile money agents in three cities, training a shared fraud-detection model without any raw transaction data leaving agent premises.

The federated model matched centrally-trained baseline performance within 2 percentage points on precision and recall, while eliminating the need for cross-border data transfer — a meaningful operational advantage given the regulatory environment mobile money operators face.

References

1. McMahan B. et al. Communication-Efficient Learning of Deep Networks from Decentralized Data. AISTATS, 2017.
2. GSMA. State of the Industry Report on Mobile Money, 2025.