FedCK: Exploiting Client Similarity for Effective Personalized Federated Learning

Federated Learning (FL) is a learning paradigm that collaboratively trains machine learning models among distributed clients while preserving data privacy. The prevalence of data heterogeneity problem underscores the need for effective personalized federated learning algorithms. However, many exiting personalized federated learning methods overlook the utilization of clients similarities. In this paper, we propose FedCK which leverages class scores to identify analogous clients and then incorporates knowledge distillation loss to transfer knowledge from average classifiers to local classifiers. Extensive experiments on EMNIST and CIFARIO dataset validate the superiority of FedCK over other FL methods.