Multimodal Mamba Fusion for Noise-Robust EEG Biometric Identification

Tóm tắt

Brain-derived signals offer a spoof resistant biometric modality, but electroencephalogram (EEG) authentication remains brittle under environmental noise and stricter held-out-subject protocols. Building on our two-stage WaveNet+ResNet pipeline, we introduce a multimodal Mamba augmented framework that (i) augments a WaveNet denoiser with a selective state-space (Mamba) block, (ii) adds a parallel spectrogram branch that performs frequency- and time-axis Mamba sweeps, and (iii) fuses the two views with a gated self-attention fusion module trained with an embedding-level mixer gate. We perform a systematic four-version comparison (V1: WaveNet only; V2: WaveNet+Mamba; V3: tuned Mamba quick run; V4: full multimodal with gated self-attention fusion) on the PhysioNet Auditory-Evoked-Potential dataset under a held-out-subject open-set protocol with session-disjoint closed set retrieval (holdout subjects {2,5,7,12}, three random seeds, three noise families). The proposed multimodal model attains the best Precision@1 on all three noise families (82.4% Gaussian, 87.7% power-line, 83.8% EMG); the corresponding Stage I reconstruction quality is SI-SNR 12.15/32.38/13.83dB. We also expose a verification gap (AUROC ≈ 0.50) on held-out subjects that motivates calibration as future work.