Clustered Federated Learning Framework with Acceleration Based on Data Similarity

Federated Learning is a distributed machine learning framework which allows multiple participants training machine learning model without exchanging their local data. It addresses critical issues such as data privacy in distributed machine learning. In real circumstances, the statistical heterogeneity of data on different devices will cause bad performance of training process. In this paper, we propose FedCSA, a clustered federated learning framework with acceleration algorithm using the similarity of data distribution between federated learning clients. Clients with similar data distribution are clustered and the acceleration algorithm is performed among them to obtain group model, which can maximum the utilization of similarity. The global model is aggregated from group models in each round of training. The empirical evaluation shows that FedCSA outperforms state-of-art approaches on datasets with different non-IID settings.