FedSemiSelf: A Hybrid Semi-Self-Supervised Federated Learning Framework

In this paper, we propose FedSemiSelf, a novel hybrid federated learning framework that addresses the challenges of limited labeled data and statistical heterogeneity in federated settings by combining self and semi supervised learning approaches. It integrates a contrastive learning objective to build robust local representations without requiring labels, while employing a confidence-based pseudo-labeling strategy that progressively introduces unlabeled data based on their estimated reliability. A curriculum scheduler dynamically controls training stages by evaluating model consistency across clients, enabling a stable and effective transition between self-supervised and pseudo-supervised phases. To enhance memory efficiency and representation diversity, each client maintains a coreset of representative samples drawn from local data streams. FedSemiSelf demonstrates strong potential for real-world decentralized learning scenarios such as smart sensing, edge intelligence, and privacy-preserving computer vision.