Distributed Federated Learning with Byzantine Resilience via Secure Multi-Party Computation

Abstract

Federated learning (FL) offers a promising approach to training machine learning models on decentralized data sources without directly exchanging data, thereby preserving user privacy. However, FL systems are vulnerable to Byzantine attacks, where malicious participants can inject corrupted model updates to compromise the global model. This paper proposes a novel framework for distributed federated learning with Byzantine resilience leveraging Secure Multi-Party Computation (SMPC). Our approach utilizes SMPC protocols to ensure that individual model updates remain private during aggregation, effectively mitigating the impact of Byzantine attacks. We formulate the problem as a multi-party computation where each participant contributes to the model update without revealing their individual data or intermediate computations. The core claim of this work is that achieving robust federated learning in the presence of malicious participants necessitates advanced privacy-preserving techniques, and our SMPC-based framework provides a compelling solution. We demonstrate the theoretical resilience of our framework against Byzantine attacks and outline the key components involved in implementing this system. The goal is to provide a secure and robust method for collaborative model training in environments with untrusted participants.