Federated learning is currently the most advanced technology in secure multi-party machine learning and has attracted significant attention for its ability to enable collaborative model training across multiple nodes without requiring data to leave local devices. However, in real-world scenarios, client data are often non-independent and identically distributed (Non-IID), which significantly degrades the performance of the global model. Current mainstream methods primarily focus on correcting model parameter biases during training, but their effectiveness diminishes as data distribution becomes more extreme. Another category of methods aims to correct data distribution biases before local training, but this approach poses privacy risks. This study proposes a framework that adheres to strict federated learning privacy protection requirements and utilizes techniques such as differential privacy and secure multi-party computation. Assuming that client data distributions remain black boxes, the most beneficial labels for improving global model performance are selected. High-quality synthetic data are generated for these labels and distributed to clients, thus making the local data distribution closer to IID and significantly enhancing the performance of the global model. Specifically, a privacy-preserving algorithm is first designed to capture the distribution of labels across clients, and clients are grouped according to these label distribution characteristics. Then, the most worthwhile labels for generation are selected, and within the corresponding groups, clients are coordinated to collaboratively train high-quality global generative models, which are stored on the server. During the training phase of the task model, based on label distribution characteristics, samples are selectively synthesized using these generative models and distributed to clients, resulting in a more uniform local data distribution. This process reduces local model disparities and yields a high-quality global model after aggregation. Experimental results demonstrate that the proposed method can effectively improve the accuracy of the global model on the test set and reduce the number of federated learning communication rounds required before convergence. Its effectiveness surpasses that of baseline methods and has been validated on different datasets.