Architecting for Privacy and Performance: A Federated Learning Framework for On-Device Financial AI in Mobile Applications

User concerns about mobile financial AI are addressed by integrating personalized services, fraud detection, and robust security. However, concerns about data centralization, the GDPR, and the global data protection streamlining make the design difficult. In this regard, this paper presents DPFed-MobileFin. This innovative cross-device Federated Learning (FL) framework specially focuses on iOS banking applications. DPFed-MobileFin combines state-of-the-art privacy-preserving methods —‘Local Differential Privacy’ (LDP) and ‘Secure Aggregation’ (SA) — with on-device ML training using Core ML and TensorFlow Lite, and an on-device implementation of Differential Privacy (DP). In addition, the system uses a Personalized FL (PFL) framework to address performance degradation caused by non-IID, heterogeneous financial data. The system architecture is developed to achieve hardware acceleration via Core ML’s Neural Engine, provide cryptographic privacy guarantees, and mitigate computational overhead. The framework demonstrates its promise for privacy and usefulness by attaining a fraud detection rate of 95.1%. The low mobile overhead, low latency, and energy usage justify the mobile user experience. Such a comprehensive system architecture meets data-residency requirements and sets a new standard for safe, highly effective AI in cross-device financial applications and systems.