Bilateral Improvement in Local Personalization and Global Generalization in Federated Learning

Federated learning (FL) is a machine learning paradigm where a server trains a global model by amalgamating contributions from multiple clients, without accessing personal client data directly. Personalized FL (PFL), a specific subset of this domain, shifts focus from a global model to providing personalized models for each client. This difference in training objectives signifies that while conventional FL aims for optimal generalization at the server level, PFL focuses on the client-side model personalization. Often, achieving both generalization and personalization in a model is challenging. In response, we introduce FedCACS, a classifier aggregation with cosine similarity in the FL method to bridge the gap between the conventional FL and PFL. On the one hand, FedCACS adopts cosine similarity and a new PFL training strategy, which enhances the personalization ability of the local model on the client and enables the model to learn more compact image representation. On the other hand, FedCACS uses a classifier aggregation module to aggregate personalized classifiers from each client to restore the generalization ability of the global model. Experiments on the public data sets affirm the effectiveness of FedCACS in personalization, generalization ability, and fast adaptation.