Exploitation Maximization of Unlabeled Data for Federated Semi-Supervised Learning

Federated learning (FL), as an emerging paradigm in edge intelligence, enables numerous edge devices to collaborate with a central server in training a shared model without compromising data privacy. Existing FL algorithms mostly focus on addressing the problem of non-independently and identically distributed (Non-IID) data in supervised scenarios. However, in practical settings, a considerable amount of unlabeled data exists, posing an urgent challenge of how to fully utilize such data. In this paper, we propose an exploitation maximization method of unlabeled data for federated semi-supervised learning (FSSL). Specifically, we adjust the optimization objective of high-confidence samples by considering both the predicted results and the class distribution of pseudo-labels. This adjustment enables the resolution of severe class imbalance issues. Additionally, we propose a diversity sampling strategy for low-confidence samples. The strategy aims to increase the diversity of training samples by sampling samples that are less similar to high-confidence samples, thereby improving the generalization ability of the model. Extensive experiments demonstrate that our method not only outperforms other state-of-the-art FSSL methods in non-IID scenarios but also exhibits superior scalability and robustness.