Dynamic Differential Privacy for Federated Learning

Abstract

This paper presents a novel approach to differential privacy in federated learning, termed Dynamic Differential Privacy (DDP). The core idea is to adapt the privacy budget dynamically, responding to the varying sensitivity of local data. Traditional federated learning frequently utilizes a static privacy budget, which can lead to either overly conservative privacy protection, significantly impacting model accuracy, or insufficient protection when confronted with malicious adversaries. DDP addresses these limitations by iteratively adjusting the privacy budget (ε, δ) based on gradient variance and update divergence. The algorithm reduces the privacy budget when gradients exhibit high variance, indicating sensitive data, and increases it when a node provides substantially different updates, signaling potential adversarial behavior. We demonstrate that this dynamic adaptation provides a stronger privacy guarantee while maintaining acceptable accuracy levels compared to fixed-budget approaches. This work contributes to a more robust and practical implementation of federated learning in privacy-sensitive environments.