Epilepsy affects 2% of the world’s population and epilepsy detection is critical for avoiding seizures. However, these data involve patient privacy and are difficult to share, hindering centralized model training. Therefore, in order to improve the accuracy of epilepsy detection while protecting patient privacy, an epilepsy detection method for EEG data based on model pruning federated learning is proposed. First, after completing the training task, the client needs to use the model pruning strategy to sparse the uploaded model parameters, and then the server uses Euclidean distance and gradient constraints to identify potential abnormal clients. Finally, the weighted aggregation method is used to assign different weights to the performance metrics of each client to improve the overall performance of the global model. Experimental results show that the accuracy of the method for seizure detection reaches 97.3%, and even if the gradient sparsity reaches 98%, the accuracy of the model decreases by only 2.7%. This approach successfully balances data privacy and model performance, and has potential in medical practice and deployment.