Domain-Tailored Attention Transformer for EEG-Based Person Identification

Accurate person identification (PI) based on electroencephalogram (EEG) provides high security value. However, existing EEG-based PI methods still struggle to fully mine individual-specific features within each do main and to effectively jointly model multi-domain information, which limits identification performance. To address these limitations, we propose the Domain-Tailored Attention Transformer (DTAT) model, which designs tailored attention modules for each domain to thoroughly mine individual-specific features, thereby enabling effective joint modeling across multiple domains. Specifically, in the spatial domain, to address the neglect of high-frequency spatial variations, we introduce a Discrete Cosine Trans form (DCT)-informed Spatial Attention (DSA) module. DSA transforms spatial EEG maps into the frequency domain and aggregates full-band components from global structures to local textures, enabling fine-grained characterization of individual-specific spatial patterns. For the temporal domain, to remedy the oversight of salient amplitude variations, we propose an Amplitude-aware Temporal Attention (ATA) module that jointly models global activation trends and localized amplitude fluctuations, adaptively emphasizing time steps with prominent amplitude responses to encode identity-critical transient temporal sig natures. To overcome the overlooking of uneven frequency band contributions in spectral modeling, we develop an Adaptive Spectral Attention (ASA) module that models localized spectral dependencies and dynamically recalibrates band-wise importance, thereby selectively enhancing identity-relevant spectral features. These three domain tailored attention modules work collaboratively to fully extract individual-specific representations from EEG signals thereby improving PI performance. Extensive experiments show the superiority of DTAT over state-of-the-art methods on multiple EEG datasets. Code is available at https://anonymous.4open.science/r/DTAT-79D6.