Privacy-Preserving Federated Learning with Differential Privacy: Methods, Threats, and Future Directions

Abstract

Federated Learning (FL) and Differential Privacy (DP) have become fundamental techniques for building privacy-preserving and decentralized machine learning systems. FL supports collaborative training of models over distributed devices and organizations without exchanging raw data and DP enhances privacy guarantees by adding mathematically calibrated noise to model updates, which prevents sensitive information to be inference by an unauthorized party. The survey provides a thorough review of the foundations, integration, security issues and real-world applications of DP-enhanced Federated Learning. It analyses major architectural frameworks of FL, fundamental concepts of DP and the ways of distributing noise to gradients or model parameters in distributed setting. Threat models that are reviewed in the paper include inference attacks, reconstruction attacks, as well as, poisoning attacks, and highlighting the significance of a robust defense and secure aggregation. Practical application in healthcare, finance, IoT, smart cities and education show the general applicability of DP-FL and benchmark analysis shows the trade-offs between privacy, utility and communication efficiency. The survey concludes with the discussion of the open challenges and future research directions which involve the adaptive privacy mechanisms, enhanced robustness, and standard evaluation frameworks. Altogether, DP-FL represents one of the most promising approaches for developing secure, reliable, and ethically responsible AI systems.