Distributed Learning with Federated Byzantine Fault Tolerance (FBT)

Abstract

This paper proposes a novel approach to distributed learning that leverages the principles of Federated Byzantine Fault Tolerance (FBT) to achieve robustness and scalability against malicious actors. Traditional federated learning systems are vulnerable to attacks where adversaries can manipulate the global model by injecting poisoned updates. Our method integrates FBT protocols, originally designed for blockchain consensus, into the distributed learning framework. This allows nodes to reach agreement on model updates even when some nodes are compromised and actively attempting to disrupt the learning process. We demonstrate that this combination provides a significant improvement in security and reliability compared to standard federated learning, offering a pathway to secure and trustworthy distributed model training. The core claim of this work is that robust and scalable distributed learning in the presence of malicious actors is a significant challenge, and our proposed solution directly addresses this issue.