K-PFed: Communication-efficient Personalized Federated Clustering

Federated Learning (FL) has attracted much attention in recent years as a promising framework of privacy-preserving and scalability. However, the framework is highly vulnerable to data heterogeneity, which widely exists in real-world applications. In this paper, we designed a communication-efficient personalized federated clustering algorithm k-PFed, which overcomes the data heterogeneity and various network scenarios in the FL system. By initializing local datasets by pre-clustering, clients in k-PFed are able to generate a personalized model locally. In this manner, the local and the global model in k-PFed can be trained independently. Meanwhile, the k-PFed designed three working modes for offline, P2P, and cloud-collaborative networks. Experiments on homogeneous and heterogeneous data show that the clustering accuracy is significantly better than other federated clustering algorithms, and the decrease caused by data heterogeneity on k-PFed is significantly lower than that of other algorithms.