BPFL: Blockchain-based privacy-preserving federated learning against poisoning attack

Yanli Ren, Mingqi Hu, Zhe Yang, Guorui Feng, Xinpeng Zhang

Abstract: Highlights•A blockchain-based privacy-preserving federated learning against poisoning attack is proposed, where nodes cooperate to detect poisonous data by computing the cosine similarity between local gradient and global gradient.•An incentive mechanism is constructed to monitor the behavior of nodes, where blockchain can promptly punish malicious nodes by confiscating their deposit submitted to blockchain before the learning task begins.•The privacy of local gradients of clients and aggregation gradient are preserved in our protocol even if multiple nodes are malicious or collusive.