Tailoring Privacy: Performance Analysis of Differential Privacy for Personalized Federated Learning (video)

This webinar session will explore personalized federated learning (PFL), a cutting-edge method that creates customized models for diverse client needs, enhancing model convergence through meta-learning. We'll address the significant challenges associated with information leakage and unveil a differential privacy (DP) based PFL framework. Our approach designs a privacy budget allocation scheme rooted in the Rényi Differential Privacy composition theory. Throughout the session, we will discuss the formulation of convergence bounds applicable to both convex and non-convex loss functions. This analysis leads to optimal strategies for determining the most effective model size and achieving the best balance among communication rounds, convergence performance, and privacy considerations. Supported by thorough evaluations on various real-life datasets, our findings corroborate our theoretical predictions and serve as a practical guide for developing DP-PFL algorithms. This webinar is perfect for researchers, data scientists, and practitioners eager to deepen their knowledge and enhance the privacy and efficiency of their federated learning projects. Join us to gain valuable insights into how to design PFL systems and ensure robust privacy protections while maintaining high performance.