Jia-Ping Lin1, Yong-Sheng Chen1,2??, and Li-Fen Chen3
1 Inst. of Biomedical Engineering, National Chiao Tung University, Hsinchu, Taiwan
2 Dept. of Computer Science, National Chiao Tung University, Hsinchu, Taiwan
3 Inst. of Brain Science, National Yang-Ming University, Taipei, Taiwan
Abstract. In this work we utilize the inter-subject differences in the electroencephalographic (EEG) signals evoked by visual stimuli for per-son identification. The identification procedure is divided into classifica-tion and verificaclassifica-tion phases. During the classificaclassifica-tion phase, we extract the representative information from the EEG signals of each subject and construct a many-to-one classifier. The best-matching candidate is further confirmed in the verification phase by using a binary classifier specialized to the targeted candidate. According to our experiments in which 18 subjects were recruited, the proposed method can achieve 96.4%
accuracy of person identification.
1 Introduction
Conventional person identification methods include passwords, smart cards, and a variety of biometric techniques. Passwords and smart cards are widely-used because of the advantage of convenience. However, smart cards might be stolen, simple passwords might be deciphered, and complicated passwords might be for-gotten. Current biometric features such as iris, fingerprints, face, voice, palm, and gait do not suffer the above-mentioned disadvantages, but they can be stolen, duplicated, or even provided under violent threats. Brainwave is an emerging bio-metric feature for person identification because of its uniqueness and consistency.
Moreover, brainwave is difficult to steal or duplicate and the characteristics em-bedded in the brainwave when the subject is under threat are hardly the same as those in normal situation. These advantages promote brainwaves as new keys to safer person identification systems.
Among all the non-invasive brainwave acquisition modalities, electroencepha-lography (EEG) has the advantages of portability, easy operation, high tempo-ral resolution, and low costs. To evaluate the uniqueness and consistency of the
?This work was supported in part by the MOE ATU program, Taiwan National Science Council under Grant Numbers: NSC-99-2628-E-009-088 and NSC-100-2220-E-009-059, and the UST-UCSD International Center of Excellence in Advanced Bio-engineering sponsored by the Taiwan National Science Council I-RiCE Program under Grant Number: NSC-99-2911-I-009-101.
?? Corresponding author.
2 Person Identification using Electroencephalographic Signals
characteristics in EEG signal, the work in [6] confirmed that the inter-subject variation of EEG spectra where different subjects administered the same task was larger than the intra-subject variation where the EEG signals of the same subject were repeatedly acquired for several times. At first resting data was used for person recognition and the identification rate ranged from 72 to 85% [10].
In 2003, Palaniappan and Ravi investigated the task-related EEG signals. By extracting features from visual evoked potentials (VEPs), the identification ac-curacy was improved to be larger than 90% [9]. The features in EEG signals include autoregressive (AR) coefficients, coherence, and cross-correlation [7]. In [1] the event-related potentials (ERPs) were utilized for person identification.
This work used the images of self-relevant objects as the visual stimuli and se-lected prominent channels related to this experiment. Temporal domain features such as P100, N170, and N250 were used in the signal analysis [4]. For simplicity and practicability, the work [5] classified subjects simply by thresholding the EEG power spectrum.
In this paper we present a person identification system using EEG signals.
Because resting state is prone to be more fluctuating, we adopt task-related EEG signals evoked by visual stimuli in this work. Representative information is extracted from the EEG signals of subjects and are used to train a many-to-one classifier for person classification. The best-matching candidate of each classifi-cation is further verified by using a binary classifier to exclude the intruder.
2 Materials
2.1 Participants and paradigm
Eighteen subjects participated in this study (age ranges from 21 to 33 years with mean 24 years, twelve males). All the subjects have normal or corrected-to-normal visions. For five participants among all the subjects, EEG data were acquired two times with an interval of more than one week.
The paradigm of data acquisition in this study is shown in Fig. 1. The subject was seated comfortably in a silent room and was asked to watch a monitor screen. The visual stimulus, an image containing either a small disk or a large one (five times larger than the small one), was presented for one second followed by another second of fixation image of a cross. The frequency ratio between the stimulus images is one (large disk) to three (small disk). Around 250 trials were acquired for each participant.
2.2 EEG recording and preprocessing
Thirty-two standard scalp electrodes were placed according to the International 10-20 System of Electrode Placement. We picked the channels related to the visual stimuli and P300 component in the frontal, frontal-central, parietal, and occipital regions [3]. The ten channels we selected were Fz, FCz, Cz, CPz, P3, Pz, P4, O1, Oz, and O2. This process will reduce the quantity of data and eliminate
Person Identification using Electroencephalographic Signals 3
+
1 s 1 s
or
Fig. 1: Paradigm for data acquisition in this study. A trial consists of one-second stimulus, an image containing either a small disk or a large one, and one-second fixation.
the activities which are not induced by the events. The EEG data were recorded with Scan 4.3 software and the sampling rate for data acquisition was 500Hz.
The earlobe electrodes A1 and A2 provided the reference. Signals were digitally filtered within the 5-30 Hz band.
We used EEGLAB 9.0 [2] to perform the following signal preprocessing pro-cedure. The EEG data were first segmented into epochs starting from one second before the stimulus onset to one second after stimulus onset. The baseline cor-rection was applied to remove the DC drift. Epochs with burst activities during the post-stimulus period were rejected (with the threshold values -50µV and 50µV). The trials evoked by the large disk events were used in the following person identification analysis.
3 Methods
3.1 Feature Extraction
For each of the EEG channels, we applied a series techniques to extract features.
These techniques, described in the following, include dimension reduction, mor-phological operation, power spectrum, and stochastic modeling.
Dimension reduction Principal component analysis (PCA) is a method for reducing feature dimension. Its main idea is to find a set of basis, usually with a much smaller dimension, to represent the original data set while preserving as much as information measured by the variance of data distribution. If there is an embedded non-linear manifold lying in a high-dimensional space and the dimension of the manifold is relatively low, this manifold can be well represented in a low-dimensional space [8]. Therefore, we also applied the locally linear em-bedding (LLE) method to transform the data to a low-dimensional space while
4 Person Identification using Electroencephalographic Signals
maintaining the manifold structure manifested in the original high-dimensional space. Firstly, we find a set of nearest neighbors for each data point Xi in D-dimensional Euclidean space. Then we reconstruct, or represent, each data point by a linear combination of its neighbors Xij with weightings Wij as the contri-bution of the neighbor Xij to this linear combination for Xi. The reconstruction error is:
E(W ) =X
i
|Xi−X
j
WijXij|2 , (1)
where the sum of the weightings for each data point Xi equals one. The data point Xi can be mapped to the corresponding point Yi in a low-dimensional space as:
Yi=X
j
WijYij , (2)
where the point Yij is the point in low-dimensional space corresponding to Xij in the original high-dimensional space.
Morphological features The latency and amplitude of each EEG epoch were computed as the morphologic features which contain VEPs (with the time inter-val from 50 ms to 150 ms after stimulus onset) and ERPs (with the time interinter-val from 250 ms to 400 ms after stimulus onset).
Frequency features The discrete Fourier transform (DFT) were used to com-pute the power spectrum for each epoch. In this work we focus on the frequency band from 5 Hz to 30 Hz.
Stochastic modeling Considering the EEG signal as an autoregressive (AR) process, we used the Yule-Walker equations to estimate the AR coefficients as the features. To fit a p th-order AR model to the EEG data X(t), we minimize the following prediction error by using the least squares regression:
X(t) =
P
X
i=1
a(i)X(t − i) + e(t) , (3)
where a(i) are the auto regression coefficients, e(t) represents the white noise, and the time series can be estimated by a linear differential equation.
Time-frequency model The wavelet transform uses a set of time-scale basis to represent the original signal. Here we applied the Daubechies wavelets to transform the time-domain EEG signals and obtained 250 coefficients as the time-frequency features.
Person Identification using Electroencephalographic Signals 5
3.2 Classification
For classification, we employed the support vector machine (SVM) and the k-nearest neighbor (kNN) search method (k=9) as the classifier. To fairly evaluate the accuracy of classification, we apply the 8-fold cross validation that separate EEG data into training and testing data to obtain the average classification accuracy for person identification.
3.3 Verification
The purpose of the verification procedure is to reconfirm the best-matching re-sult of classification. For each of the eighteen subjects, we trained a SVM binary classifier by using two groups of training data including EEG data of the targeted subject and those of all others. We evaluate the binary classifier for verification in terms of the true acceptance rate (TAR) and the false acceptance rate (FAR).
The best-matching subject from the classification procedure is verified by the corresponding binary classifier. In addition, we modified the false classified data in classification phase through iterative verification. The probability estimate, which is a confidence level of classification, determines an ordered list of can-didates having confidence levels larger than 80% of that of the best-matching candidate.
4 Results
4.1 Temporal characteristics in the acquired signals
We first verified whether the resting EEG or ERP is better for distinguishing subjects’ identities. By applying the SVM classifier to categorize the pre-stimulus (500 ms before onset) EEG signals among the eighteen subjects, the classification accuracy was 12.2%. When the post-stimulus (500 ms after onset) ERP signals were used for person identification, the classification accuracy achieved 25.3%.
Therefore the ERP contains more information for person identification than resting EEG does.
4.2 Accuracy in the classification phase
Table 1 shows the classification accuracy comparison among seven features extracted from the 1000ms post-stimulus EEG signals with respect to single trial, average of two trials, SVM, and kNN. The average of two trials can achieve higher classification accuracy compared to single trial data because of higher signal-to-noise ratio. Regarding the classifier, SVM outperforms kNN with respect to various features.
Among the seven kinds of features, power spectrum achieves the best clas-sification accuracy while the latency and amplitude generally lead to poor re-sults. Fig. 2 shows the power spectrum of different subjects with the frequency band ranging from 5 Hz to 30 Hz. We can see that within-subject variation of
6 Person Identification using Electroencephalographic Signals
Table 1: Results of classification with different features and different classifiers.
The data of each subject acquired in the same experiment.
SVM kNN
Feature Single trial Avg (2 trials) Single trial Avg (2 trials)
Raw data 29.31% 80.86% 23.47% 76.38%
LLE 30.81% 86.69% 28.13% 83.44%
PCA 27.74% 83.48% 25.32% 81.28%
Latency 11.59% 35.23% 10.21% 33.56%
Amplitude 38.53% 50.82% 36.23% 45.19%
Power spectrum 72.03% 91.61% 60.01% 85.92%
AR 53.52% 62.54% 50.96% 60.57%
Wavelet 27.27% 85.41% 22.92% 77.26%
Table 2: TAR in the verification phase, which is the percentage of the best-matching candidates in the classification phase that are accepted in the verifica-tion phase.
Subject 1 2 3 4 5 6 7 8 9
TAR (%) 97.14 98.70 98.81 100 100 96.43 97.62 100 96.30
Subject 10 11 12 13 14 15 16 17 18
TAR (%) 97.96 100 100 98.57 100 100 100 100 100
the spectra of different trials is smaller than inter-subject variation. In order to accommodate different information of the best two features, we combined the power spectrum and LLE features after normalization and achieve 97.1% of classification accuracy.
4.3 Accuracy in the verification phase
The true acceptance rate (TAR) measures the percentage of the best-matching candidates in classification that are accepted by the binary classifier of verifica-tion. Table 2 shows the TARs of eighteen subjects and their average is 97.9%.
The false acceptance rate (FAR) is zero, that means all the false classified data were successfully rejected in the verification phase. After iterative verification the overall accuracy of our system is 96.4%.
4.4 Classification accuracy over time
For five participants among all the eighteen subjects, EEG data were acquired two times with an interval of more than one week. The goal is to verify whether the EEG data of the same subject is sufficiently stable for person identification over a period of time. The average accuracy of the classification phase is 93.2%.
Table 3 shows the TAR, FAR and results of iterative verification. After iterative verification, the overall identification accuracy of our system is 85.7%, indicating
Person Identification using Electroencephalographic Signals 7
5 10 15 20 25 30
0 1 2 3 4
Subject1
Frequency (Hz)
Power Spectrum
5 10 15 20 25 30
0 2 4 6
Subject2
Frequency (Hz)
Power Spectrum
Fig. 2: The power spectrum of ten trials of two subjects (thick black lines rep-resent the averages of ten trials). Each trial shows the average results of ten channels.
Table 3: TAR, FAR, and results of iterative verification of data acquired from different days.
Subject 3 5 8 12 13
TAR (%) 78.57 100 97.92 83.33 100 Accepted/False classified 0/1 4/6 0/0 0/0 11/12
FAR (%) 0 66.67 - - 91.67
that the the performance of our system slightly degrades over time. One possible remedy is to retrain the classifier by adding the data acquired over time so that the classifier can be adapted to each subject. By using the two sets of EEG data acquired at different times, the average accuracy of the classification phase is improved from 93.2% to 98.4%, TAR is increased from 90.6% to 97.7%, and FAR is decreased from 79.0% to 0%. After iterative verification, the overall identification accuracy of our system is improved from 85.7% to 96.8%.
5 Discussion and Conclusions
The major causes affecting the accuracy of person identification using EEG sig-nals include both external and internal interferences. The external interferences deteriorate the quality of acquired signal whereas the internal interferences re-sult in signal instability over time. From the calculated correlation between EEG trials of different subjects, the EEG data of subjects having high correlation to those of other subjects have more classification errors. Compared with the inter-subject correlation, the intra-inter-subject correlation between EEG trials acquired
8 Person Identification using Electroencephalographic Signals
at different times is higher. Therefore, the brainwave signals are suitable for biometric measures for person identification.
We have proposed a person identification system using visual-evoked EEG signals. According to our experiments, we concluded that the combination of power spectrum and LLE can extract informative features for distinguishing subjects. The identification system contains the classification and verification phases. In the classification phase, we use a multi-class classifier to perform a one-to-many comparison for each acquired data. In the iterative verification phase, the best-matching candidates are furthered verified sequentially by binary classifiers according to their matching levels. The overall person identification accuracy of the proposed system can achieve 96.4%.
References
1. Berlad, I. and Pratt, H.: P300 in response to the subject’s own name. Electroen-cephalogr Clin Neurophysiol. vol. 96, pp. 472-4. Ireland(1995)
2. Delorme, A. and Makeig, S.: EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. Journal of Neuro-science Methods. vol. 134, pp. 9-21. (2004)
3. Donchin, E. and Coles, M. G. H. Is the P300 Component a Manifestation of Context Updating. Behavioral and Brain Sciences. vol. 11, pp. 357-374(1988)
4. Miyakoshi, M. and Nomura, M. and Ohira, H. An ERP study on self-relevant object recognition. Brain Cogn. vol. 63, pp. 182-9(2007)
5. Miyamoto, C. and Baba, S. and Nakanishi, I. Biometric Person Authentication Using New Spectral Features of Electroencephalogram (EEG). Ispacs 2008. pp. 130-133(2008)
6. Napflin, M. and Wildi, M. and Sarnthein, J. Test-retest reliability of EEG spectra during a working memory task. NeuroImage. vol. 43, pp. 687-93(2008)
7. Riera, A. and Soria-Frisch, A. and Caparrini, M. and Grau, C. and Ruffini, G. Un-obtrusive biometric system based on electroencephalogram analysis. Eurasip Journal on Advances in Signal Processing. (2008)
8. Roweis, S. T. and Saul, L. K. Nonlinear dimensionality reduction by locally linear embedding. Science. vol. 290, pp. 2323-6(2000)
9. Singhal, G. K. and RamKumar, P. Person identification using evoked potentials and peak matching. Biometrics Symposium. pp. 156-161(2007)
10. Poulos, M. and Rangoussi, M. Person identification from the EEG using nonlinear signal classification. Methods of Information in Medicine. vol. 41, pp. 64-75(2002)