Differential Privacy for Federated Learning via Secure Multi-Party Computation

Abstract

Federated learning (FL) presents a promising paradigm for training machine learning models on decentralized data sources without directly exchanging the data itself. However, achieving differential privacy (DP) – a rigorous privacy guarantee – within the FL setting remains a significant challenge due to the computational overhead associated with traditional DP mechanisms. This work introduces a novel approach that combines differential privacy with secure multi-party computation (SMPC) to address this limitation. Our method employs a layered SMPC protocol to enable clients to perform gradient updates locally while maintaining privacy. The protocol minimizes data exposure by breaking down the computation into smaller, secure steps. The core claim is that existing DP mechanisms in FL are often computationally expensive. The proposed mechanism combines differential privacy with secure multi-party computation (SMPC) to perform gradient updates locally without revealing individual client data, utilizing a layered SMPC protocol for enhanced efficiency. This offers a practical and efficient solution for deploying DP in FL. The theoretical analysis demonstrates that our approach can achieve a desired privacy budget while significantly reducing the computational burden compared to standard DP techniques. This work contributes a new method to improve the performance of DP in FL.