Federated Semi-Supervised Learning (FSSL) is a distributed learning framework designed to leverage limited labeled data alongside abundant unlabeled data on client devices, while preserving data privacy. In FSSL, clients utilize unlabeled data via data augmentation and consistency regularization. However, some samples, after undergoing strong augmentation, achieve near-zero loss due to their high confidence classification, indicating that they are well-learned and contribute little to further model optimization. These samples are termed “original samples”. Existing FSSL methods fail to address the challenge posed by original samples, applying a one-size-fits-all strategy that diminishes their potential contribution. FedSatch uniquely addresses this gap. For the purpose of solving the dilemma caused by this situation, we propose FedSatch, a framework that dynamically adapts to original samples. FedSatch consists of two key modules: (1) a client-side adaptive sample identification and augmentation module, which uses Exponential Moving Average (EMA) and OTSU thresholding to distinguish between original and routine samples, applying specialized augmentations to the former, and (2) a global federated optimization mechanism built on the FedAvg algorithm, which minimizes model drift and enhances training efficiency through cross-client parameter aggregation and periodic global synchronization. Extensive experiments on benchmark datasets such as CIFAR-10 and SVHN demonstrate that FedSatch significantly improves accuracy, with gains of 11.60% and 16.48%, respectively, over existing FSSL methods. FedSatch presents a scalable, robust solution for real-world FSSL tasks, addressing the challenge of imbalanced sample informativeness while maintaining data privacy in distributed environments.