Graph-regularized federated learning with shareable side information

This study focuses on specifying local models in federated learning (FL), which allows a large number of clients to improve their corresponding models by training a shared global model. However, current FL models often fail to consider the difference between the data distributions in various clients while enforcing all local models to be identical, thus leading to a considerable loss of local personalization. To this end, this study proposes a graph-regularized federated learning framework, GraphFL, by exploiting the available client features commonly shared with other clients in the real world. Specifically, GraphFL achieves the similarity matrix of all clients using the permitted shareable side information and subsequently updates local models by returning a specific model from the server instead of an identical model. The proposed model iteratively learns the neural network parameters for each client. Compared with state-of-the-art FL models, GraphFL can benefit from the employed similarity and achieve improved classification performance in clients on three publicly available image datasets.