Learning latent neural signatures from EEG scalograms for robust biometric identification

Abstract noindentBiometric identification using electroencephalography (EEG) offers a promising avenue for developing robust identification systems. However, achieving high accuracy across a large population remains a challenge due to the low signal-to-noise ratio and high inter-subject variability of EEG signals. This paper proposes a novel methodology that converts 1D EEG signal into 2D cortical latent scelogram maps. More specifically, 64-channel EEG epochs are first transformed into wavelet scalograms and compressed into latent representations using a convolutional autoencoder. These latent features are then spatially arranged onto a grid according to the scalp electrode layout for further classification. For biometric identification, resting state EEG activity is used to train a custom Residual Neural Network (ResNet) for extracting subject-specific patterns. Evaluated on the PhysioNet Motor Imagery dataset comprising 109 subjects, the proposed system employs a session-level majority voting strategy to mitigate single-epoch noise. The proposed approach achieves robust performance, attaining 99.1% accuracy in subject classification that significantly outperforms single-segment baseline methods and demonstrating strong scalability and suitability for reliable EEG-based biometric identification.