Federated Learning (FL) faces challenges with Non-Independently and Identically Distributed (Non-IID) client data. To fundamentally address this issue, we propose a FedDM: Federated Semi-Supervised Learning (FSSL) method based on Diffusion Model, which investigates FSSL from the perspective of data augmentation. Specifically, we train a diffusion model on clients using unlabeled data and aggregate its parameters to the server. The server then generates a synthetic unlabeled dataset conforming to clients’ class distributions, which is distributed to local clients in an IID manner to mitigate the Non-IID data challenge. The local clients select unlabeled data using Class-Balanced Adaptive Threshold and select a balanced dataset from the synthesized data on the server for balance training the global model. Experimental results demonstrate that our method outperforms state-of-the-art approaches on three natural datasets. The code is available at https://github.com/liangguirong/FedDM .