Federated learning is a distributed machine learning paradigm that allows multiple participants to collaborate on training a shared model without revealing their own data. However, federated learning faces challenges such as data heterogeneity, communication overhead, security attacks, etc., which result in low efficiency of data training. To address these issues, this paper proposes a strategy to enhance the training efficiency of federated learning data through hyperparameter optimization and proxy re-encryption. A semi-supervised model of federated learning based on generative adversarial network is designed so that each participant can automatically adjust and optimize their own training parameters according to their own data distribution and environment in order to adapt to the data heterogeneity and dynamics of federated learning. Finally, experiments demonstrate that the proposed method can significantly improve the efficiency of data training and model performance compared with traditional federated learning methods across different datasets and scenarios. Experimental results show that the proposed strategy can achieve the same or even better accuracy than centralized machine learning while safeguarding data privacy and enabling ciphertext conversion.