CBRFL: A framework for Committee-based Byzantine-Resilient Federated Learning

Abstract

Federated Learning (FL), a decentralized machine learning paradigm, has gained attention for enabling collaborative model training without sharing raw data. However, traditional FL architectures rely on a central server, creating trust issues, single points of failure, and vulnerabilities to Byzantine attacks due to the lack of effective gradient validation. In this paper, we introduce the Committee-Based Byzantine-Resilient Federated Learning Framework (CBRFL), which decentralizes using a blockchain-based off-chain committee consensus mechanism for gradient validation and adaptive aggregation, eliminating the need for a central server. Furthermore, we present a momentum and adaptive global learning rate mechanism to improve training stability, along with a contribution and reputation system to enhance the reliability of committee members. The experimental results show that CBRFL outperforms robust FL algorithms across four federated heterogeneous datasets and three attack methods. Without attacks, CBRFL performs similarly to leading heterogeneous FL baselines in most scenarios.