A Privacy-Preserving Hybrid Deep Learning Framework for Real-Time Fraud Detection in the Ugandan Mobile Money Ecosystem

Mobile money offers a widely used channel for payments and financial services in Uganda, but fraud detection must address limited labels, changing behaviour, transaction speed and the protection of personal data. This article develops a proposed privacy-preserving hybrid deep learning framework from the supplied literature review. It synthesises research on rule-based and machine learning approaches, sequential and graph-based models, federated learning, differential privacy and streaming evaluation. A narrative review was undertaken using the supplied chapter, targeted SciSpace searches and verification against research publications and Ugandan legislation. The literature suggests that sequential models can represent patterns across transactions and that collaborative training may be possible without pooling raw records. Neither capability establishes that a hybrid model will outperform simpler approaches in Uganda. Federated updates can disclose information unless additional protections are designed and evaluated, while differential privacy provides a defined, parameterised guarantee rather than absolute secrecy. The proposed architecture separates collaborative model training from local transaction scoring and retains human review and customer safeguards. Evaluation should use time-ordered data, realistic baselines, precision-recall measures, decision latency, falsepositive costs, privacy accounting and subgroup analysis. No model was trained and no Ugandan transaction dataset was analysed for this article. The findings are therefore a synthesis of prior evidence and a testable design, not measured detection results. A controlled pilot with authorised providers and independent privacy and operational assessment is needed before deployment.