Development of an EEG-Based Biometric Identification System Using a Hybrid CNN-LSTM Architecture and Electrode Attention Mechanism

Abstract

Electroencephalography (EEG) signals offer a promising, spoof-resistant modality for continuous biometric Identification systems. However, their practical implementation is hindered by low signal-to-noise ratios, computational complexity, and the high cost of dense medical-grade devices. To overcome these fundamental challenges, this study proposes a novel, hybrid deep learning architecture (CNN-LSTM) enhanced by a spatial electrode attention mechanism, specifically designed to operate efficiently on low-cost, 14-channel consumer-grade wearable devices. The methodology relied on extracting "eyes-closed" resting-state signals from 21 participants, strictly applying a temporal data splitting protocol to prevent complete data leakage. The system effectively combines convolutional layers for spatial feature extraction with recurrent networks for temporal sequence modeling. Experimental results demonstrated the model's superior capability in decoding nonlinear spatio-temporal patterns, achieving an exceptional accuracy of 99.71% on strictly independent temporal test data, and an average accuracy of 99.79% across temporal cross-validation folds. Furthermore, the proposed model features a remarkably lightweight footprint with under 20,000 trainable parameters, making it highly ideal for low-power computations and continuous real-time biometric Identification applications.