Epileptic seizure detection from electroencephalography signals currently represents a critical clinical task that demands both high accuracy and explainability, while also adhering to strict privacy regulations. We propose a privacy-preserving and explainable federated learning-based method for epileptic seizure detection that combines two wavelet-based representations with a Vision Transformer classifier. The electroencephalography signals are first converted into audio recordings and then transformed into time–frequency representations using the Morlet and Mexican Hat continuous wavelet transforms, generating two datasets. Local models are trained on distributed clients using a Vision Transformer as a feature extractor, and global knowledge is aggregated via Federated Averaging and Federated Proximal algorithms. To address the lack of transparency, we employ Attention Rollout to highlight regions most relevant to the model’s predictions. Experimental results on two wavelet datasets achieve interesting performance, by reaching 0.817 of accuracy, thus showing the effectiveness of the proposed method.