EEG Signal Analysis for Human Verification using Neural Networks – Preliminary Experimental Results

Abstract

The results of experimental studies on human verification by EEG signal analysis are presented in this paper. The developed approach was investigated using 220 EEG examinations recorded from 11 people, 20 examinations for every person. The first fifteen examinations were used for neural networks learning, and the rest 5 examinations for their evaluation. The EEG signals recorded for every person were separated into short segments for which feature extraction was conducted. After that, the features were introduced to a feedforward neural network, trained by the Levenberg-Marquardt backpropagation algorithm. We focused on spectral features, calculated separately for four EEG frequency bands. After the network training, we evaluated our approach by introducing at the network inputs the examinations from other days that were not presented to the neural network before. The results for two electrode sets: placed on the central (C3, Cz, C4, C3CzC4) and centro-occipital (C3, C4, O1, O2, C3C4, O1O2, C3C4O1O2), using accuracy, sensitivity, specificity, and precision measures, are presented and discussed in this paper. Regardless of the number of electrodes, almost all mean metrics were above 0.70 and increased with the number of electrodes from which the EEG signal features fed the neural network. One of the aims of this work was to investigate, whether temporary, daily changes in EEG signals would prevent people from being recognized.