CR-IFSSL: Imbalanced Federated Semi-Supervised Learning with Class Rebalancing

Federated semi-supervised learning is a popular research area known for its ability to preserve data privacy, utilize client data efficiently, and reduce the need for a large number of manually labeled samples. However, most existing studies focus on scenarios where client data is independently and identically distributed (IID). In practical applications, client data often deviates from the IID assumption, posing significant challenges for existing federated semi-supervised learning algorithms. In this paper, we proposed a novel approach called Class Rebalancing Imbalanced Federated Semi-Supervised Learning (CR-IFSSL) designed specifically to address the class imbalance issue in non-IID settings, thereby expanding the applicability of federated semi-supervised learning. The primary principle of our method is to conduct local class rebalance training for each individual client. Specifically, to handle the problem of class imbalance, each client’s contribution is weighted according to its estimated class distribution when selecting pseudo-label samples. This ensures that samples from minority classes are chosen more frequently, thereby mitigating the class imbalance issue. Furthermore, our method gradually adjusts the alignment strength in the selftrained predictive data distribution to fine-tune the model’sperformance. Experimental results demonstrate that CR-IFSSL outperforms existing federated semi-supervised learning algorithms when dealing with class imbalance data, thereby offering valuable insights for real-world data applications.