Securing Distributed SGD Against Gradient Leakage Threats

This paper presents a holistic approach to gradient leakage resilient distributed Stochastic Gradient Descent (SGD). First , we analyze two types of strategies for privacy-enhanced federated learning: (i) pruning with random selection or low-rank filtering and (ii) perturbation additive noise differential privacy noise. We the inherent limitations these approaches their underlying impact on guarantee, model accuracy, attack resilience. xmlns:xlink="http://www.w3.org/1999/xlink">Next present securing SGD in learning, controlled as tool. Unlike conventional methods per-client injection fixed parameter strategy, our keeps track trend per-example updates. It makes adaptive closely aligned throughout training. xmlns:xlink="http://www.w3.org/1999/xlink">Finally provide an empirical analysis utility, resilience proposed approach. Extensive evaluation using five benchmark datasets demonstrates that can outperform state-of-the-art competitive accuracy performance, strong high against attacks.