Privacy-Preserving Federated Learning Framework via Blockchain and Committee Mechanism.

Federated learning allows multiple clients to cooperatively train an efficient model without exposing clients' local data. However, this distributed training technique is susceptible to Byzantine attack that obstruct the training of global model by altering the model or uploading erroneous gradients. To solve this issue, in this paper, we propose an novel committee based Byzantine-tolerant federated learning algorithm (CBTFL), which guarantees the resilience, convergence, and correctness of training. A committee is established in CBTFL to review the uploaded local gradients from the clients. The committee rates the uploaded gradients to the clients using the scoring system, and the server qualifies and aggregates the gradients based on the scoring results by utilizing the qualified function, and at the end of each training round, the members in committee are switched by the selection strategy. Besides, we employ the Cheon-Kim-Kim-Song (CKKS) algorithm, a fully homomorphic encryption scheme, and blockchain to encrypt and record the training process with the goal of maintaining clients' privacy. According to the experiment results, CBTFL is more robust than conventional Byzantine-tolerant algorithms.