When Differential Privacy Meets Wireless Federated Learning: An Improved Analysis for Privacy and Convergence

This paper studies how to safely train AI models across multiple devices (federated learning) while protecting user privacy using differential privacy. The key contributions are showing that privacy loss can stay bounded rather than grow indefinitely during training, and providing theoretical guarantees for model performance with realistic assumptions like gradient clipping. The work addresses practical concerns in wireless federated learning with non-convex optimization, moving beyond oversimplified prior analyses.