pFSSL-D: Generalization Meets Personalization in Dual-Phase Federated Semi-Supervised Learning

Federated Semi-Supervised Learning (FSSL) offers a distributed learning paradigm that addresses the critical issue of label scarcity while preserving client privacy. However, current FSSL methods are often hindered by an over-reliance on labeled data for initialization, high communication overhead, and suboptimal global model performance in heterogeneous data settings. To overcome these limitations, we propose pFSSL-D, a novel Dual-Phase Generalization and Personalization Pipeline designed to generate several models for unlabeled clients. In the first phase, decentralized contrastive learning with feature alignment is proposed to efficiently pre-train a robust and generalizable feature extraction model while minimizing communication overhead. In the personalization phase, we introduce a parameter-granularity federated fine-tuning approach with semantic consistency, which decouples model parameters into general and personalized components, providing specialized update strategies for each component. This method effectively balances global generalization with client-specific adaptation, ensuring robustness in heterogeneous environments. Extensive evaluations on benchmark datasets show that pFSSL-D consistently outperforms state-of-the-art FSSL methods in terms of accuracy, convergence speed, and resource overhead.