Multidimensional fusion EEG-based identity recognition using DAF-Net

Electroencephalography (EEG) signals originate from individual neural activity and possess inherent advantages such as stealthiness, resistance to spoofing, revocability, and intrinsic liveness, making them a promising modality for high-security identity authentication. However, EEG signals are highly non-stationary and characterized by low signal-to-noise ratios, while identity-discriminative information is distributed across multidimensional structures, including time–frequency dynamics and spatial topological patterns. More importantly, differences in neural activation mechanisms induced by distinct experimental paradigms introduce substantial distributional variation, causing many existing methods to perform well within a single paradigm while showing limited robustness under heterogeneous paradigm settings. To address this issue, we aim to improve the stability of EEG-based identity recognition under heterogeneous paradigm conditions and propose a DAF-Net-based framework that combines dual-view generative augmentation with adaptive fusion. Specifically, the same EEG signal is represented from two complementary views—a time–frequency representation (TFR) and a topographic map (Topomap)—and view-specific generative augmentation is performed using 2 K-cGAN and FastGAN-ASP to alleviate data scarcity and noise perturbations. A heterogeneous dual-branch feature extraction network is then designed, and cross-attention is employed to explicitly model the complementary interactions between the two views, enabling mutual guidance and feature enhancement. Finally, an adaptive weighting strategy based on Bayesian principles is applied to obtain a fused representation for identity classification. Experiments conducted on the motor imagery BCI Competition IV-2b dataset and the naturalistic deception Bag-of-Lies dataset demonstrate that the proposed method achieves consistently strong identification performance across two substantially different paradigms. Attention-based visualizations further validate the complementarity between time–frequency and spatial information, suggesting that the proposed framework provides improved robustness and stability under substantially different task paradigms.