This paper aims to shed light on how EEG-based end-to-end deep learning models process time-series raw EEG signals in order to generate predictions. The paper uses frequency domains because of 2 advantages: the strong correlation with cognitive states and the inherent capacity to model long-range temporal dependencies. The paper aims to address the gap in using this perspective in research through FourierMask. FourierMask is the first mask perturbation framework designed specifically for frequency-domain explanation of EEG-based end-to-end models. Through this method, three key innovations are introduced. First, the Fourier-based domain transformation enables direct manipulation of spectral components. Then, a learnable mask mechanism jointly models the spectral-spatial couplings relationship to explain EEGs. Finally, a perturbation generator consigned by a target alignment loss ensures natural perturbations by minimizing distribution shift via cluster-aware regularization. The methods were validated by experimenting on an EEG benchmark dataset using EEGNet, TSCeption, and DeepConvNet models. The method reached a 36.0% average accuracy drop gap (compared to 8.6% for LIME and 6.6 for easy PEASI) at the group level. It also reaches a 17.8% average accuracy drop gap at the instance level compared to 8.9% for LIME and 9.9% for easyPEASI.