Exploring One-shot Semi-supervised Federated Learning with A Pre-trained Diffusion Model

Federated learning is a privacy-preserving collaborative learning approach. Recently, some studies have proposed the semi-supervised federated learning setting to handle the commonly seen real-world scenarios with labeled data on the server and unlabeled data on the clients. However, existing methods still face challenges such as high communication costs, training pressure on the client devices, and distribution differences among the server and the clients. In this paper, we introduce the powerful pre-trained diffusion models into federated learning and propose FedDISC, a Federated Diffusion Inspired Semi-supervised Co-training method, to address these challenges. Specifically, we first extract prototypes from the labeled data on the server and send them to the clients. The clients then use these prototypes to predict pseudo-labels of the local data, and compute the cluster centroids and domain-specific features to represent their personalized distributions. After adding noise, the clients send these features and their corresponding pseudo-labels back to the server, which uses a pre-trained diffusion model to conditionally generate pseudo-samples complying with the client distributions and train an aggregated model on them. Our method does not require local training and only involves forward inference on the clients. Our extensive experiments on DomainNet, Openimage, and NICO++ demonstrate that the proposed FedDISC method effectively addresses the one-shot semi-supervised problem on Non-IID clients and outperforms the compared SOTA methods. We also demonstrate through visualization that it is of neglectable possibility for FedDISC to leak privacy-sensitive information of the clients.