PFedSim: An Efficient Federated Control Method for Clustered Training

To further promote the implementation and application of artificial intelligence technology under data silos, federated ecology has defined a set of standardized and systematic solutions from upstream data to downstream applications from a systematic perspective. As the core of the federated ecology, federated control can transform the abstract tasks of federated management into concrete execution rules. Clustered training is a common cooperation rule for the personalized federated training task of multi-party cooperation. However, it is difficult to generalize in large-scale cooperation scenarios because of increased data transmission pressure and server storage burden. Therefore, this paper designs a personalized federated control method based on similarity measures(pFedSim) so that the server can accurately distribute the cluster model to clients, thereby reducing the transmission cost. Compared with the existing clustered training method IFCA, pFedSim transfers the selection rights of cluster models from the client to the server. Model selection preference, accurate distribution of cluster models. Experiments show that in the personalized training task with
${s}$
clusters, the method can obtain the same model level with 1/s data transmission. It is better than the current optimal clustering training method Macro-Acc increased by 2.25%.