A Decentralized Federated Learning Framework via Committee Mechanism With Convergence Guarantee

Federated learning allows multiple participants to collaboratively train an efficient model without exposing data privacy. However, this distributed machine training method is prone attacks from Byzantine clients, which interfere with the of global by modifying or uploading false gradient. In article, we propose a novel serverless federated framework Committee Mechanism based Learning (CMFL), can ensure robustness algorithm convergence guarantee. CMFL, committee system set up screen uploaded local gradients. The selects gradients rated elected members for aggregation procedure through selection strategy, and replaces member election strategy. Based on different considerations performance defense, two opposite strategies are designed sake both accuracy robustness. Extensive experiments illustrate that CMFL achieves faster better than typical Learning, in meanwhile obtaining traditional Byzantine-tolerant algorithms, manner decentralized approach. addition, theoretically analyze prove under strategies, coincides experimental results.