<p>Federated Semi - Supervised Learning (FSSL) Federated Semi-Supervised Learning (FSSL) addresses data silos by enabling collaborative training across decentralized clients. However, non-independent and identically distributed (non-IID) data and imbalanced model reliability among clients often degrade global model performance. To tackle these challenges, we propose a random consensus model integrated with pseudo-label similarity matching. Each client model is treated as a biased estimator, and a random subsampling strategy is employed to extract consensus representations. Aggregation weights are adaptively adjusted based on the quality of sub-consensus models, improving robustness against heterogeneous data distributions. Furthermore, a voting mechanism over time enhances the quality of pseudo-labels by integrating global model outputs, thereby improving unlabeled data utilization through refined similarity matching. Experimental results on multiple datasets demonstrate the effectiveness of our approach compared to existing FSSL methods.</p>