On the potential of EEG for biometrics: combining power spectral density with a statistical test

Abstract

The objective of this work was to explore the potential of using subject's electroencephalogram (EEG) as a biometric identifier. EEG was collected from eight healthy male participants, while exposing them to the sequence of images displayed on the screen. The averaged, over EEG rhythms, estimates of power spectral density were used as the classification features for the artificial neural network and Euclidean distance-based classifiers. Prior the classification, Kruskal-Wallis test was performed on the power estimates to verify that they were statistically different between different individuals, who were performing identical tasks. Assuming the significance level of 0.075, Kruskal-Wallis analysis indicated that up to 96.42% of such estimates were statistically different between different participants and, therefore, can be used as the classification features for biometric authentication. When using average EEG spectral power as the classification features, the highest classification accuracy of 87.5% was achieved for α1 EEG rhythm (8–10 Hz), while using the artificial neural network classifier, and for α2 EEG rhythm (10–14 Hz), while using the Euclidean Distance classifier. The classification performance may be mediated by the type of visual stimulation (i.e., the image the subject perceives) and the statistical test may be instrumental for classification feature selection.