Ensemble Knowledge Distillation for Federated Semi-Supervised Image Classification

Abstract

Federated learning is an emerging privacy-preserving distributed learning paradigm, in which many clients collaboratively train a shared global model under the orchestration of a remote server. Most current works on federated learning have focused on fully supervised learning settings, assuming that all the data are annotated with ground-truth labels. However, this work considers a more realistic and challenging setting, Federated Semi-Supervised Learning (FSSL), where clients have a large amount of unlabeled data and only the server hosts a small number of labeled samples. How to reasonably utilize the server-side labeled data and the client-side unlabeled data is the core challenge in this setting. In this paper, we propose a new FSSL algorithm for image classification based on consistency regularization and ensemble knowledge distillation, called EKDFSSL. Our algorithm uses the global model as the teacher in consistency regularization methods to enhance both the accuracy and stability of client-side unsupervised learning on unlabeled data. Besides, we introduce an additional ensemble knowledge distillation loss to mitigate model overfitting during server-side retraining on labeled data. Extensive experiments on several image classification datasets show that our EKDFSSL outperforms current baseline methods.