PrSeFL: Achieving Practical Privacy and Robustness in Blockchain-Based Federated Learning

With the help of artificial intelligence, the large amount of data generated by IoT has unleashed significant value. Federated learning is emerging as a novel paradigm which can be applied to solve the privacy issues caused by analyzing IoT data. However, traditional federated learning protocols are vulnerable to inference and poisoning attacks. Various solutions have been proposed to enhance data privacy and robustness. Nonetheless, most of these solutions are usually centralized and rely on unrealistic security assumptions. Furthermore, the recently proposed blockchain-based decentralized solutions generally incur high costs, which is unaffordable for resource-constrained IoT devices. In this paper, we propose a practical secure federated learning system named PrSeFL. We utilize blockchain to decentralize federated learning process so that the security assumptions are easier to achieve in practice. To preserve data privacy, we implement secure multiparty computation based secure aggregation in blockchain environment. To guarantee practical robustness, we enforce norm constraints on the masked updates via zero-knowledge proof. Moreover, we propose a modified dynamic accumulator which is utilized to realize lightweight anonymous authentication of users. Simulation results show that, compared with state-of-the-art systems, PrSeFL has superior performance on authentication and model training. And the advantage of PrSeFL becomes more significant as the number of users grows.