Decentralized Federated Learning via Byzantine-Resilient Graph Protocols

Abstract

Federated learning (FL) has emerged as a promising approach to training machine learning models on decentralized data sources without directly exchanging data. However, the inherent vulnerability of current FL protocols to Byzantine attacks – where malicious nodes inject false information to corrupt the global model – poses a significant threat to data privacy and model accuracy. This paper proposes a novel decentralized federated learning protocol leveraging a Byzantine-resilient graph protocol. Our approach utilizes a distributed graph structure where nodes communicate and learn through this graph, incorporating mechanisms to detect and mitigate the influence of malicious actors. The core claim is that current FL protocols are susceptible to Byzantine attacks. The proposed mechanism introduces a robust decentralized system capable of handling adversarial behavior, thus improving the overall resilience of the learning process. We introduce a framework for designing and implementing such protocols, focusing on graph construction, message authentication, and consensus mechanisms. The resulting system offers enhanced data privacy and model accuracy compared to traditional, centralized FL methods.