DARN: A Dual Attention Refinement Network for Enhancing Feature Robustness in VEP-Based EEG Biometrics by Honggang Liu & Han Yang & Dongjun Liu & Hangjie Yi & Bingfeng He & Yong Peng & Wanzeng Kong instant download

DARN: A Dual Attention Refinement Network for Enhancing Feature Robustness in VEP-Based EEG Biometrics by Honggang Liu & Han Yang & Dongjun Liu & Hangjie Yi & Bingfeng He & Yong Peng & Wanzeng Kong instant download

Abstract—Visual evoked potential (VEP)-based EEG biometfrom the brain’s cerebral cortex via non-invasive electrodesrics provide a secure, spoof-resistant approach for identification[1]. Unlike traditional biometric modalities such as fingerprintsand authentication; however, cross-session variability, driven byand facial features, which cannot be revoked or re-issued oncetemporal fluctuations in neural responses, often underminescompromised, EEG-based biometrics allows for re-enrollmentfeature stability and degrades performance. To tackle this, weafter data leakage, thus reducing long-term security risks [2].propose the Dual Attention Refinement Network (DARN), a novelFurthermore, this technology inherently requires live brainmethod that enhances the spatiotemporal consistency of EEGrepresentations without requiring frequent retraining. DARNactivity, providing robustness against spoofing attacks [3].combines a lightweight CNN backbone with two complemenIn recent years, despite significant progress in various EEGtary attention modules: the Spatial Feature Refinement Unitparadigms, the low signal-to-noise ratio (SNR) of EEG signals(SFRU), which prioritizes consistent spatial patterns, and theoften necessitates prolonged data collection to achieve highInter-channel Refinement Unit (ICRU), which captures stableaccuracy. For example, resting-state EEG (eyes open or closed)inter-channel dependencies, jointly refining the spatial and channel dimensions of extracted EEG feature maps. Evaluated onmay require 60 to 90 seconds [4], [5], auditory evokedtwo public multi-session VEP datasets with 30 and 54 subjects,potentials (AEP) around 30 seconds [6], and motor imagerywith sample durations of 6 seconds for the 30-class dataset and(MI) about 15 seconds [7]. Such prolonged acquisition times4 seconds for the 54-class dataset, DARN surpasses state-ofseverely hinder the usability of EEG bi