EEG-based biometric identification with convolutional neural network

Although more interest arising in biometric identification with electroencephalogram (EEG) signals, there is still a lack of simple and robust models that can be applied in real applications. This work proposes a new convolutional neural network with global spatial and local temporal filter called (GSLT-CNN), which works directly with raw EEG data, not requiring the need for engineering features. We investigate the performance of the GSLT-CNN model on datasets of 157 subjects collected from 4 different experiments that measure endogenous brain states (driving fatigue and emotion) as well as time-locked artificially induced brain responses such as rapid serial visual response (RSVP). We evaluate the GSLT-CNN model against the comparable SVM, Bagging Tree and LDA models with effective feature selection method. The results show the GSLT-CNN model is highly efficient and robust in training more than 279 K epochs within less than 0.5 h and achieves 96% accuracy in identifying 157 subjects, which is 3% better than the best accuracy of SVM on selected PSD feature, 10% better than that of SVM on selected AR feature and 23% better than that of normal CV-CNN model on raw EEG feature. It demonstrates the potential of deep learning solutions for real-life EEG-based biometric identification. We also show that the cross-session identification accuracy from time-locked RSVP data (99%) is slightly higher than that from single-session non-time-locked driving fatigue data (97%) and much higher than that from epochs measuring random brain states (90%), which implies RSVP could be a more beneficial design to achieve high identification accuracy with EEG and our GSLT-CNN model is robust for cross-session identification in RSVP experiment.