Blockchain-empowered secure federated learning system: Architecture and applications

【Author】 Yu, Feng; Lin, Hui; Wang, Xiaoding; Yassine, Abdussalam; Hossain, M. Shamim

【Source】COMPUTER COMMUNICATIONS

【影响因子】5.047

【Abstract】Federated learning (FL) is a promising paradigm to realize distributed machine learning on heterogeneous clients without exposing their private data. However, there is the risk of single point failure with FL because it relies on a central server to gather the model updates from clients, moreover, malicious behaviors of some clients may lead to low-quality or even poisoned global models. Blockchain as a revolutionary distributed ledger technology can alleviate the above problems to significantly enhance the security and scalability of FL systems. Therefore, this article presents a general framework of Blockchain-based Federated Learning (BFL) system with detailed description of its key technologies and operation steps. We then review and compare the most recent representative BFL applications. And we outlook some key challenges and opportunities of the future BFL system in terms of security, cost, and scalability. Finally, we propose PoS-BFL in IoT scenarios with malicious devices. The validator voting mechanism and role switching mechanism in PoS-BFL ensure the stakes of legitimate nodes, and effectively reduce the impact of malicious nodes on the accuracy of the system model. And the experiments are conducted to demonstrate that PoS-BFL can achieve 86% accuracy, which is much higher than vanilla FL and pFedMe, and PoS-BFL is robust to some extent by adjusting the ratio of workers, validators and miners.

【Keywords】Blockchain; Federated learning; Deep learning; Internet of Things; Intelligent transportation

【发表时间】2022 DEC 1

【收录时间】2023-01-14

【文献类型】实验仿真

【主题类别】

区块链技术-协同技术-联邦学习