BED: A New Data Set for EEG-Based Biometrics

Abstract

ernuEqonz¡ lezD lo nd utsiginnisD tmos nd erevlilloErerr¡ ezD wiguel nd mznD xeem @PHPIA 9fihX e new dtset for iiqEsed iometrisF9D siii snternet of things journlF F

Single-channel EEG-based subject identification using visual stimuli

“…The BED dataset [9] was selected for studying the performance of individual EEG channels for the task of biometric subject identification. BED is a recently released dataset designed for EEG-based biometrics using low-cost consumergrade EEG devices.…”

“…Baseline results for subject identification were provided for BED in [9], where the EEG signals were segmented into segments of 5 s length with 50% overlapping, pre-processed to remove noise and artefacts, and MFCC features, Autoregression Reflection Coefficients (ARRC) features, and Spectral features (spectral centroid, spectral bandwidth, spectral crest factor, and spectral flatness) were computed for each segment. The problem of subject identification was then modelled as a multi-class classification problem where each class referred to an individual subject (21 classes).…”

“…Baseline results showed that for all types of features, the highest accuracies were achieved for the "Resting with eyes closed" stimulus, reaching 47.79% for the MFCC features. When only visual stimuli are considered, the highest classification accuracies were achieved for the "Images with emotional content" stimulus, reaching 40.25% for the MFCC features [9].…”

“…In this work, contrary to [9], to examine the performance of individual EEG channels, each model was trained and tested using features extracted from each individual channel, resulting in the creation of 14 multi-class models for each of the three features and each of the nine visual stimuli. Similar to [9], the EEG signals were pre-processed using the EEGLAB toolbox [11] to apply the PREP pipeline [12], consisting of line-noise removal via filtering, referencing the signal relative to the estimate of the "true" average reference, and finally detecting and interpolating bad channels relative to the reference. Then, similar to the baseline BED experiment [9], the MFCC features were computed by first applying the Fourier Transform, followed by a filterbank in the Mel scale, and the Discrete Cosine Transform.…”

“…To address these drawbacks, a public benchmark dataset called "Biometric EEG Dataset" (BED) [9] was recently proposed for EEG-based biometrics. The BED dataset contains EEG recordings from 21 individuals that were acquired during three separate sessions, each one week apart.…”