COSMOS: Model-Agnostic Personalized Federated Learning with Clustered Server Models and Pseudo-Label-Only Communication
COSMOS is a model-agnostic framework for personalized federated learning that allows clients to use any local model without revealing its architecture or gradients. It achieves this by having clients generate pseudo-labels on a shared dataset, which the server uses to cluster clients and train specialized teacher models for each group. This method provides a theoretical guarantee of exponential risk contraction for personalization.