A Study of the Effects of Batch Sampling in Semi-Supervised Federated Learning

This paper explores the implementation of batch sampling techniques in semi-supervised federated learning (SSFL), particularly focusing on environments where labeled data is scarce and primarily resides on the server side, while clients handle predominantly unlabeled data. We investigate three batch sampling methods—Uniform, Complementary, and Enforcing—within the SemiFL framework. They are tested under a practical system environment where very few labeled data are available and the clients' class distributions are distributed in non-independent and identically distributed (Non-IID) settings. Our experimental results offer insights into the impacts of these sampling strategies on learning efficacy. The findings indicate that tailored batch sampling can slightly enhance model accuracy, particularly in scenarios with limited labeled data, suggesting a potentially useful direction for further research in semi-supervised federated learning paradigms.