Federated learning (FL) can learn models from decentralized data in a privacy-preserving way by exchanging model updates instead of raw data. Most of the existing FL methods only utilize decentralized labeled data for model training. However, in many real-world scenarios, the labeled data on local clients may be rather limited due to the high cost of annotation, and most of local data may remain unlabeled. Effectively exploiting the decentralized unlabeled data has the potential to improve the performance of federated learning with limited labeled data. However, directly applying semi-supervised learning to federated learning may not achieve optimal performance, due to the well-known non-IID problem in federated learning. In this paper, we propose a semi-supervised federated learning method via adaptive switched training, named FedAST, which can effectively incorporate unlabeled local data into federated model learning. In FedAST, each client first updates the global model on local labeled data in a supervised way, and then updates it on the local unlabeled data via self-training. In order to handle the non-IID problem, on server we maintain two global models to encode both supervised and semi-supervised information, and adaptively select the optimal one for local model training on clients. Extensive experiments on two benchmark datasets show FedAST can significantly improve the performance of federated learning by effectively exploiting the unlabeled local data.