Dual Class-Aware Contrastive Federated Semi-Supervised Learning

Federated semi-supervised learning (FSSL), facilitates labeled clients and unlabeled clients jointly training a global model without sharing private data. Existing FSSL methods mostly focus on pseudo-labeling and consistency regularization to leverage the knowledge of unlabeled data, which have achieved substantial success on raw data utilization. However, their training procedures suffer from the large deviation from local models of labeled clients and unlabeled clients and the confirmation bias induced by noisy pseudo labels, which seriously damage the performance of the global model. In this paper, we propose a novel FSSL method, named Dual Class-aware Contrastive Federated Semi-Supervised Learning (DCCFSSL), which considers the local class-aware distribution of individual client's data and the global class-aware distribution of all clients' data simultaneously in the feature space. By introducing a dual class-aware contrastive module, DCCFSSL builds a common training goal for different clients to reduce the large deviation and introduces contrastive information in the feature space to alleviate the confirmation bias. Meanwhile, DCCFSSL presents an authentication-reweighted aggregation method to enhance the robustness of the server's aggregation. Extensive experiments demonstrate that DCCFSSL not only outperforms state-of-the-art methods on three benchmarked datasets, but also surpasses the FedAvg with relabeled unlabeled clients on CIFAR-10 and CIFAR-100 datasets. To our best knowledge, we are the first to present the FSSL method that utilizes only 10\% labeled clients of all clients to achieve better performance than the standard federated supervised learning that uses all clients with labeled data.