Biometric Identification based on an EEG Signal Using an Attention-enhanced Bidirectional GRU Network

Reliable identification of individuals from physiological signals is an active research direction in biometrics. Electroencephalography (EEG) is one such modality because brain signals carry individual-specific patterns. This study investigates whether a deep learning–based EEG identification framework can achieve reliable closed-set individual identification on a 109-subject benchmark under a standard subject-dependent evaluation protocol. Verification scenarios and the associated metrics (FAR, FRR, EER) are out of scope and are not reported. This work proposes and validates an attention-augmented Bi-GRU model for EEG-based individual identification, aiming to achieve high identification accuracy on a large multi-subject public dataset.

Time-domain features were extracted from 64-channel EEG recordings collected during left-hand motor movement/imagery tasks from 109 subjects. A three-layer Bi-GRU model with a self-attention mechanism was trained for a 109-class identification task using a 70/15/15% subject-dependent train–validation–test split. Training robustness was supported through training-only SMOTE-based class balancing and Gaussian-noise augmentation.

Under the adopted subject-dependent split, the proposed model achieved a mean test accuracy of 96.68% ± 0.25% across three independent training runs (95% Student-t CI 96.04–97.31%), a top-5 accuracy of 99.37% ± 0.15%, and a small validation–test gap of +0.37 ± 0.90 percentage points. Performance was higher than several traditional machine learning and simpler deep learning baselines and was comparable to reported transformer and graph neural network results.

The attention mechanism highlighted discriminative temporal patterns from EEG signals; performance under the subject-dependent split indicated that our model was able to maintain stable identification performance across the evaluated recordings.

The attention-enhanced Bi-GRU is able to capture discriminative temporal EEG patterns and shows strong identification performance under the adopted protocol, supporting further investigation of EEG-based biometric identification systems.