FedTMatch: Self-Supervised Federated Semi-Supervised Learning with Dynamic Threshold

Federated semi-supervised learning (FSSL) has emerged as a promising paradigm to address the challenges of data privacy and the scarcity of labeled data. Most existing algorithms adapt the pseudo-labeling and consistency regularization methods in traditional semi-supervised learning (SSL) to exploit raw data from unlabeled clients. However, these methods always use the same high fixed threshold to filter out noisy pseudo-labels for each class, without considering the varying learning conditions and difficulty levels associated with different classes, resulting in a decline in the performance of global model, especially under non-Independent and Identically Distributed (Non-IID) settings. To this end, we propose FedTMatch, a learning framework which adjusts the threshold for each class based on the learning status of each client. Furthermore, we introduce a self-supervised rotation loss in unlabeled clients to increase the utilization of raw data and correct the feature space formed by pseudo-label data to weaken the impact of Non-IID data distribution on global model's performance. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our approach.