Low-cost electroencephalogram (EEG) based authentication

Abstract

A low-cost, consumer-grade, EEG-based individual authentication system is proposed in this work. While EEG signals are recorded, the subject performs four mental imagery tasks consisting of a baseline measurement, referential limb movement, counting, and rotation for 150 seconds each. The 150 seconds of data are divided into one second segments, from which features are obtained. Three sets of features are extracted from each electrode: 6th order autoregressive (AR) coefficients, power spectral density, and total power in five frequency bands. Two additional sets of features are extracted from interhemispheric data: interhemispheric power differences and interhemispheric linear complexity. These feature sets are combined into a feature vector that is then used by a linear support vector machine (SVM) with cross validation for classification. The goal was to minimize both false accept rates (FARs) and false reject rates (FRRs). Using voting rules across groups of ten segments, we were able to achieve 100% classification accuracy for each subject in each task. Though more work must be done with a larger subject pool as well as across multiple sessions, these results show that low-cost EEG authentication systems may be viable.