Distributed Differential Privacy for Federated Learning

Abstract

This paper presents a novel approach to integrating differential privacy into federated learning (FL) systems. The core challenge in FL lies in protecting user data while still achieving high model accuracy. This work introduces a distributed differential privacy (DDP) mechanism specifically designed to address this challenge. Our method utilizes a local perturbation strategy at each client, combined with a global privacy accounting protocol. This approach guarantees differential privacy at the client level, minimizing the risk of individual data disclosure, without incurring a significant drop in model accuracy. We demonstrate that this tailored DDP solution offers a more sophisticated and effective approach compared to existing methods, providing a robust framework for privacy-preserving FL. The key contributions of this paper are the design of the local perturbation strategy and the global privacy accounting protocol, both optimized for the unique constraints of the federated learning setting. This results in a system that balances privacy protection and model utility.