Distributed Differential Privacy via Secure Multiparty Computation for Federated Learning

Abstract

Federated learning (FL) offers a promising approach to training machine learning models on decentralized data sources without directly exchanging data. However, this approach inherently exposes sensitive user data, raising significant privacy concerns. This paper proposes a novel framework for distributed differential privacy (DDP) within a secure multiparty computation (SMC) environment for FL. The core idea is to utilize SMC to guarantee differential privacy while minimizing communication overhead and computational complexity. Our protocol constructs a distributed computation where each client contributes to a shared global model update, shielded by the privacy guarantees of differential privacy. The proposed system aims to achieve a strong privacy-utility trade-off, addressing a critical gap in existing FL solutions. We formally define the system, outline the protocol, and analyze its privacy and computational characteristics. The key contributions lie in the synergistic combination of SMC and differential privacy within a federated setting, providing a robust solution for privacy-preserving machine learning.