FedPVD: Clustered Federated Learning with NoN-IID Data

Federated Learning (FL) is a distributed machine learning framework, which makes full use of private data of client for model training under the premise of protecting client privacy. Although the research on federated learning has made great progress, in the real world federated learning system, data samples are usually scattered and NoN-IID, which will seriously lead to the deterioration of federated learning performance. In order to solve the problem of the deterioration of federated learning performance caused by NoN-IID data, this paper proposes a clustered federated learning algorithm based on the direction of client model parameter change, which is used to improve the accuracy of federated learning model and reduce the influence of NoN-IID data on federated learning. Moreover, this algorithm supports any proportion of clients to participate in federated learning. Specifically, on the premise of not increasing the calculation amount of the client, through analyzing the changes of the model parameters uploaded by each client at the server, the changes are taken as the direction of their local update for clustering, based on this, a new aggregation method is designed. Experimental results show that the proposed algorithm is significantly better than the baseline federated learning algorithm under various NoN-IID and IID data, and has strong robustness under different client participation rates.