EEG-Based Identification Using Conformer Networks

In recent years, biometric authentication technologies such as fingerprint, iris, facial, and voice recognition have gained widespread adoption as advanced tools for secure identity verification. However, these modalities remain vulnerable to identity leakage due to limited anti-spoofing and anti-coercion capabilities. Medical research has revealed that electroencephalogram (EEG) signals generated by the human brain not only inherit conventional biometric properties but also exhibit unique advantages in spoof resistance, coercion resistance, and liveness detection. Consequently, EEG-based identification technologies have emerged, with deep learning methodologies demonstrating particular prominence in effectively extracting high-dimensional data features. This study investigates the application of the Conformer model-a deep learning architecture integrating convolutional neural network (CNN) with Transformers-to EEG-based identification. The model enhances global attention mechanisms while capturing localized features, thereby improving generalization capabilities and classification accuracy. Experiments were conducted using the SEED-VII emotion recognition dataset from Shanghai Jiao Tong University, where EEG samples evoked by emotional stimuli were employed for participant identification. The proposed model achieved an accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$98.02 \% \pm 3.4 \%$</tex>, significantly outperforming conventional approaches. For comparative analysis, baseline models including CNN and EEGNet were evaluated under identical conditions, yielding accuracies of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 0. 9 \%} \pm \mathbf{2. 9 2 \%}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 8. 5 \%} \pm \mathbf{1. 1 2 \%}$</tex>, respectively. These results robustly validate the experimental conclusions. Further analysis identified differential entropy features of EEG signals as the most critical contributor to authentication performance.