Decentralized Defense: Leveraging Blockchain against Poisoning Attacks in Federated Learning Systems

Abstract

<p>Federated learning (FL) has become the next generation of machine learning (ML) by avoiding local data sharing with a central server. While this becomes a major advantage to client-side privacy, it has a trade-off of becoming vulnerable to poisoning attacks and malicious behavior of the central server. As the decentralization of systems enhances security concerns, integrating decentralized defense for the existing FL systems has been extensively studied to eliminate the security issues of FL systems. This paper proposes a decentralized defense approach to FL systems with blockchain technology to overcome the poisoning attack without affecting the existing FL system's performance. We introduce a reliable blockchain-based FL (BCFL) architecture in two different models, namely, Centralized Aggregated BCFL (CA-BCFL) and Fully Decentralized BCFL (FD-BCFL). Both models utilize secure off-chain computations for malicious mitigation as an alternative to high-cost on-chain computations. Our comprehensive analysis shows that the proposed BCFL architectures can defend in a similar manner against poisoning attacks that compromise the aggregator. As a better measure, the paper has included an evaluation of the gas consumption of our two system models.</p>