DARN: A Dual Attention Refinement Network for Enhancing Feature Robustness in VEP-Based EEG Biometrics

Visual evoked potential (VEP)-based EEG biometrics provide a secure, spoof-resistant approach for identification and authentication; however, cross-session variability, driven by temporal fluctuations in neural responses, often undermines feature stability and degrades performance. To tackle this, w...

Full description

Saved in:
Bibliographic Details
Published inIEEE transactions on information forensics and security Vol. 20; pp. 7166 - 7180
Main Authors Liu, Honggang, Yang, Han, Liu, Dongjun, Yi, Hangjie, He, Bingfeng, Peng, Yong, Kong, Wanzeng
Format Journal Article
LanguageEnglish
Published IEEE 2025
Subjects
Accuracy
Authentication
Biometrics
Brain modeling
Data mining
Deep learning
Electroencephalogram(EEG)
Electroencephalography
Feature extraction
person identification
Robustness
visual evoked potentials
Visualization
Online AccessGet full text

Cover

More Information
Summary:Visual evoked potential (VEP)-based EEG biometrics provide a secure, spoof-resistant approach for identification and authentication; however, cross-session variability, driven by temporal fluctuations in neural responses, often undermines feature stability and degrades performance. To tackle this, we propose the Dual Attention Refinement Network (DARN), a novel method that enhances the spatiotemporal consistency of EEG representations without requiring frequent retraining. DARN combines a lightweight CNN backbone with two complementary attention modules: the Spatial Feature Refinement Unit (SFRU), which prioritizes consistent spatial patterns, and the Inter-channel Refinement Unit (ICRU), which captures stable inter-channel dependencies, jointly refining the spatial and channel dimensions of extracted EEG feature maps. Evaluated on two public multi-session VEP datasets with 30 and 54 subjects, with sample durations of 6 seconds for the 30-class dataset and 4 seconds for the 54-class dataset, DARN surpasses state-of-the-art baselines, achieving identification accuracies of 93.83% (30 classes) and 84.55% (54 classes), and authentication equal error rates of 3.05% and 3.85%, respectively. Moreover, our analysis highlights the pivotal role of visual stimulus diversity in improving cross-session generalization, offering practical insights for designing robust VEP-based biometric systems. The source code is available at https://github.com/Ultramua/DARN .
ISSN:1556-6013
1556-6021
DOI:10.1109/TIFS.2025.3587181