Power Spectral Density Analysis for Human EEG- based Biometric Identification

Power Spectral Density Analysis for Human EEGbased Biometric Identification Zhi Ying.Ong Biosignal Processing Research Group (BioSIM) PPK Mechatronic, Universiti Malaysia Perlis Arau, Perlis, Malaysia jean-624@hotmail.com A.Saidatul Biosignal Processing Research Group (BioSIM) PPK Mechatronic, Universiti Malaysia Perlis Arau, Perlis, Malaysia saidatul@unimap.edu.my Abstract—Authentication is most important for security. There are many different systems for recognizing the person. The traditional authentication systems such as passwords have drawbacks. It is easy to be stolen. Biometric authentication systems provide the best security. However, a current technique that widely used for identification which is fingerprint has its own disadvantages. Furthermore, current techniques such as facial recognition, iris recognition and voice recognition that used to recognize person still compromise the security walls. In this recent years, electroencephalograph (EEG) signal has been discovered that it has the potential to become one of the biometric authentication systems. It is brain activities for a human. It is unique due to the EEG signal is different from person to person. In this paper, power spectral density analysis was used to analyse the electroencephalography (EEG) signal. Knearest neighbor classifier was used for classification in this paper. The accuracy results of alpha (8 – 13 Hz), beta (13- 30 Hz), combined alpha and beta (8 – 30 Hz) and combined theta, alpha, beta and gamma (4 – 40 Hz) frequency bands were compared. Overall, the percentage of accuracy was above 80%. The most suitable frequency bands for human EEG-based biometric identification in this experiment was the combined theta, alpha, beta, and gamma (4 – 40 Hz). The percentage of accuracy for this frequency band was the highest among the others which is 89.21%. Keywords—biometric, EEG, identification, recognition, PSD I. INTRODUCTION Human identification is recognizing a person. Commonly, we make use of passwords to access social networks, e-mail, even bank account. Passwords for security is simple and easy but it can be easily circumvented [1]. The existence of too many password-account pairings for each user is the biggest problem. This is because the user tends to forget or use the same username and password for multiple sites [2]. The traditional methods such as passwords, PINs and RF cards are easily forgotten, stolen or lost [3]. Figure 1 shows the traditional authentication system which is by using passwords. 978-1-5386-8369-9/18/$31.00 ©2018 IEEE Z.Ibrahim Technopreneur at UniMAPSdn. Bhd. (TUSB) Universiti Malaysia Perlis Kangar, Perlis, Malaysia Fig. 1. Traditional authentication method by using passwords [4]. Biometric provides more secure authentication compared to the traditional authentication system. The word biometrics originated from Greek words which are bios = “life” and metron = “measure”. Biometric authentication has divided into two categories which are Conventional Biometrics and Cognitive Biometrics. Conventional Biometrics either use physiological characteristic which is “the way the individual possesses” or behavioural characteristic which is “the way the individual behaves”. The examples of physiological characteristic are fingerprint and iris scan. Signature and voice are the examples of behavioural characteristic. Biometric detected during cognitive or emotional brain states is known as Cognitive Biometrics. It is based on the measurement of signals directly or indirectly produced by “the way the individual thinks”[5]. The biometric authentication system that widely used is fingerprint system. However, approximately 1-4% of the population unable to use their fingerprint [6], [7]. There are many people unable to provide clear fingerprint images especially the patients with skin diseases. The skin disease like hand dermatitis will influence the fingerprint quality and recognition [8], [9], [10]. Figure 2 shows the picture of skin disease. Figure 3 shows the unclear fingerprint images. Electroencephalograph (EEG) are noisy and complex signals. It is easily affected by the body movements and the activities that we did even different brain activities[11].EEG signals have divided into several frequency bands [4]. Table I shows the EEG frequency bands and corresponding brain activities. Due to its noisy and complex signals, advanced signal processing must carry out to enhance the EEG signals. Three main stages for signal processing namely preprocessing, feature extraction and classification. TABLE I. FREQUENCY BANDS AND CORRESPONDING BRAIN STATES [4] FREQUENCY BAND Fig. 2. Skin disease [9]. Delta (δ) 1 – 4 HZ Theta (θ) 4 - 8 HZ Alpha (α) 8 - 13 HZ Beta (β) 13 – 30 HZ Gamma (γ) 30 – 100 HZ BRAIN STATE Primarily associated with deep sleep. Appear as consciousness slips towards drowsiness. Usually found over the occipital region. Indicates relaxed awareness without attention. Associated with active thinking and concentration. Represents binding of different populations of neurons. II. EXPERIMENTAL PROTOCOL AND METHODS Fig. 3. Unclear fingerprint images [10]. Facial recognition is one of the biometric authentication systems. However, people’s faces change over time. Facial recognition also affected by lighting or facial expressions. Iris recognition is difficult to capture. It is easily obscured by eyelids, lens and reflections from the cornea. Voice recognition required as little background noise as possible. Illness can change a person’s voice. A. Data Set and Experimental Protocol Due to the relatively stable features and simple operation, visual stimuli are commonly used in the field of biomedicine [13]. With different kind of stimuli, the respond of brain activity can be recorded as a response to those stimuli [14]. In this experiment, blue colour paper, owe identity card and other people’s identity card act as visual stimuli to trigger the EEG signals. Electroencephalograph (EEG) signals can possibly be used for building a robust and most secure biometric identification system. EEG reflect the inner self of a person and it is unique. When performing similar mental activities, the EEG signals are different from person to person. So, they are impossible to forge[11]. EEG signal is a secret mental task which cannot be observed. The EEG signals of similar mental tasks are person dependent. The brain signals are sensitive to stress and mood of the person. The EEG signals will be different if the person is under stress. Therefore, EEG signals are confidential, difficult to mimic, impossible to steal and it requires a living person to produce the signals [12]. Electroencephalograph (EEG) signals were collected from ten subjects. The Subjects were asked to visualize three things which were blue colour paper, owe identity card and other people’s identity card. Each thing visualized for 20 seconds. The EEG signals collected by using an EEG device which namely EEGO sports. It contains 32 channels and the512 Hz sampling rate. The subjects were asked to avoid body movement. This is because noise will be detected which caused by body movements due to artifacts [15], [16]. In order to reduce electrode-scalp impedance, a conductive gel was used [15]. Figure 4 shows the photo during data collection. Fig. 5. Raw signal for first channel. Fig. 4. Photo during data collection. B. EEG Signal Processing After data collection, EEG signal processing was carried out to analyse the EEG signals. Preprocessing, feature extraction and classification are the three main stages for signal processing. Preprocessing is used to remove unwanted noise and feature extraction is used to extract most useful information for human authentication. Classification is used to identify “patterns” of the extracted biometric features. 1) Preprocessing In the stage of preprocessing, Butterworth bandpass filter was designed. Butterworth filter does not have roll-off of minus 20dB per pole and passband ripple. It is defined mathematically by cut off frequency and a number of poles[17], [18]. Butterworth filter produced flat frequency response for the purpose of analysis [19]. Due to no ripple, the best compromise between attenuation and phase response is the Butterworth filter [18]. Before filter, the raw signal contains all frequency. In order to separate the signal according to the frequency bands, the raw signal has to be filtered. Signal for theta contains 4Hz to 8Hz, signal for alpha contains 8Hz to 13Hz, signal for beta contains 13Hz to 30Hz while signal for gamma contains 30Hz to 40Hz. Figure 5 shows the raw signal for the first channel. Figure 6, figure 7, figure 8 and figure 9 show the filtered signal for theta, alpha, beta and gamma respectively. The x-axis for all graphs below is amplitude and y-axis for all graphs below is time. Fig. 6. Filtered signal for theta. Fig. 7. Filtered signal for alpha. Fig. 8. Filtered signal for beta. 3) Classifier The classifier is to predict the corresponding class of the independent features as input to which an independent variable belongs [25], [26]. The classification required a training dataset and a testing dataset. The trained classifier will model the association between classes and corresponding features. It is capable of identifying new instances in an unseen testing dataset. K-fold cross-validation is used to confirm the classification models in the present work were trained and tested with EEG data. This technique is widely used. The performance of a single classifier on a given dataset will be evaluated by k-fold cross-validation. It also can compare the performances of two classification algorithms. This technique uses all instances in a dataset for either training or testing. Instance means the validation for each data is exactly once [25]. K-nearest neighbor (KNN) classifier was the method that used for classification in this experiment. It is a non-linear classifier [1]. This classifier is simple but robust. It is capable to produce high-performance results even for complex applications [27]. A testing sample’s class is identified by knearest neighbor according to the majority class of k-nearest training samples [25]. In order to determine the data belongs to which group, KNN uses a distance of features in a data set. When the distance between the data is close, a group is formed. When the distance between the data is far, many groups are formed[27]. III. RESULTS AND DISCUSSION Fig. 9. Filtered signal for gamma. 2) Feature Extraction Power spectral density (PSD) analysis was used for feature extraction. It is a well-established and commonly used method for signal processing. Power spectral density is defined as the distribution of signal power over frequency[20], [21]. It shows the strength of the energy as a function of frequency [22]. Power spectral density analysis is used for quantification of EEG signals[20], [21]. A sensitive means for detecting periodicity within the waveforms and determining the relative energy content of the periodicities are provided by this mathematical method of frequency analysis of complex waveforms which is power spectral density analysis[21], [23]. Welch’s method is one of the power spectral density analysis. It is a non-parametric method. The signal is divided and sequenced into segments. This method also offers to reduce noise [21], [23]. The signal to noise ratio for Welch’s method is high. It reduces in the estimated power spectra in exchange for reducing the frequency resolution[22], [24]. By changing the Bartlett method in two aspects, an improved estimator was obtained which is Welch’s method. The two aspects are allowed the data segments in the Welch’s method to overlap and windowed each data segment prior to computing the periodogram. It is also an improved version of the periodogram [24]. In this experiment, the frequency bands were analysed and compared for accuracy. The frequency bands were analysed in terms of all seconds. Welch’s method was used in this experiment. K-nearest neighbors (KNN) was used for classification. Class 1 is referring to the blue colour paper, class 2 is referring to own identity card and class 3 is referring to other people’s identity card. A. Alpha The frequency band for alpha (α) is 8 – 13 Hz. This frequency band indicates the person is in the relaxing situation. It usually found over the region of occipital. TABLE II. THE ACCURACY FOR ALPHA 1 82.59 PERCENTAGE OF ACCURACY (%) CLASS 3 AVERAGE 84.82 81.70 83.04 2 82.14 82.81 81.25 82.07 3 4 5 6 7 8 81.99 80.80 79.82 79.84 80.10 79.52 81.99 80.80 81.43 81.25 81.25 81.08 80.80 81.03 80.80 79.46 79.72 79.41 81.60 80.88 80.68 80.18 80.36 80.00 CLASS 1 CLASS 2 9 80.01 80.80 79.96 80.26 10 80.36 80.44 80.09 80.30 Table II showed the accuracy for alpha. Overall, the percentage of accuracy was above 80%. The maximum average percentage of accuracy for alpha was 83.04%. B. Beta The frequency band for beta (β) is 13 – 30 Hz. This frequency band indicates the person is in the thinking and concentration situation. TABLE III. D. Combined Theta, Alpha, Beta and Gamma The frequency bands for theta (θ) and gamma (γ) are 4 – 8 Hz and 30 – 100 Hz respectively. Theta, alpha, beta and gamma were combined for analysis. The frequency range is 4 – 40 Hz in this experiment. TABLE V. GAMMA PERCENTAGE OF ACCURACY (%) THE ACCURACY FOR BETA PERCENTAGE OF ACCURACY (%) CLASS 1 CLASS 2 CLASS 3 AVERAGE 1 2 3 4 5 6 7 8 9 88.84 87.50 87.80 86.71 86.52 86.16 85.59 85.21 85.37 89.29 88.62 86.46 86.05 86.16 86.01 85.84 85.44 85.62 86.16 86.38 86.76 85.94 85.80 85.12 85.27 84.88 84.82 88.10 87.50 87.00 86.24 86.16 85.76 85.57 85.17 85.17 10 85.09 85.45 85.22 85.25 Table III showed the accuracy for the beta. Overall, the percentage of accuracy was above 80%. The maximum average percentage of accuracy for alpha was 88.10%. C. Combined Alpha and Beta The alpha and beta were combined for analysis. The frequency range is 8 – 30 Hz. TABLE IV. THE ACCURACY FOR COMBINED THETA, ALPHA, BETA AND THE ACCURACY FOR COMBINED ALPHA AND BETA PERCENTAGE OF ACCURACY (%) CLASS 1 CLASS 2 CLASS 3 AVERAGE 1 87.50 87.05 85.71 86.76 2 86.50 86.61 85.49 86.20 3 86.01 85.57 85.71 85.76 4 5 6 85.99 86.16 86.24 85.55 85.63 85.57 85.49 85.04 84.19 85.68 85.61 85.33 7 8 9 86.13 85.69 85.59 85.20 85.04 85.00 84.28 84.21 84.35 85.20 84.98 84.98 10 85.45 85.00 84.55 85.00 Table IV showed the accuracy for combined alpha and beta. Overall, the percentage of accuracy was above 80%. The maximum average percentage of accuracy for alpha was 86.76%. CLASS 1 CLASS 2 CLASS 3 AVERAGE 1 2 3 4 5 6 7 87.50 86.44 87.05 87.05 86.74 86.51 86.60 89.06 88.90 88.24 87.78 87.50 87.30 87.39 91.07 88.73 87.80 87.92 87.92 87.72 87.56 89.21 88.02 87.70 87.58 87.39 87.18 87.19 8 9 10 86.92 86.87 86.96 87.47 87.30 87.04 87.60 87.59 87.31 87.33 87.25 87.11 Table V showed the accuracy of combined theta, alpha, beta and gamma. Overall, the percentage of accuracy was above 80%. The maximum of the average percentage of accuracy for alpha was 89.21%. IV. CONCLUSION Overall, the percentage of accuracy is above 80%. This proves that the proposed methods which are power spectral density, Welch’s method and K-nearest neighbors classifier able to obtain higher classification rate. In this experiment, the most suitable frequency band is combined theta, alpha, beta and gamma in all seconds due to the percentage of accuracy is the highest among others which is 89.21%. In the future, the researchers should consider the comparison study to analyse the effectiveness for the designed protocols in terms of biometry aspects. The designed acquisition protocol will affect the performance of the EEG-based biometric system. Researchers also should focus on the development of the new algorithm for EEG biometric features that will further enhance the recognition accuracy process. Fusion features between time-frequency domain analysis and nonlinear method are expected can describe the signals generated by biological systems in a more effective way. ACKNOWLEDGEMENT We would like to thank Universiti Malaysia Perlis for supporting our works. REFERENCES [1] G. Rashmi, C. Balaji, and C. R. Rashmi, “Human Emotion Classification From EEG using SVM and K-NN Classifier For BCI,” Int. J. Adv. Res. Electron. Commun. Eng., vol. 4, no. 7, pp. 1927–1930, 2015. [2] P. Peer and J. Bule, “Building Cloud-based Biometric Services,” Informatica, vol. 37, pp. 115–122, 2013. [3] Q. Gui, Z. Jin, and W. Xu, “Exploring EEG-based biometrics for user identification and authentication,” in 2014 IEEE Signal Processing in Medicine and Biology Symposium, IEEE SPMB 2014 - Proceedings, 2014. “Using Brain Waves as New Biometric Feature for Authenticating a Computer User in Real-Time,” Int. J. Biometric Bioinforma., vol. 7, no. 1, pp. 49–57, 2013. [16] M. Fatourechi, A. Bashashati, R. K. Ward, and G. E. Birch, “EMG and EOG artifacts in brain computer interface systems: A survey,” Clin. Neurophysiol., vol. 118, no. 3, pp. 480–494, 2007. [4] I. Jayarathne, M. Cohen, and S. Amarakeerthi, “BrainID: Development of an EEG-based biometric authentication system,” in 7th IEEE Annual Information Technology, Electronics and Mobile Communication Conference, IEEE IEMCON 2016, 2016. [17] M. Sandhu, S. Kaur, and J. Kaur, “A Study on Design and Implementation of Butterworth, Chebyshev and Elliptic Filter with MatLab,” Int. J. Emerg. Technol. Eng. Res., vol. 4, no. 6, pp. 111– 114, 2016. [5] K. C. Reshmi, P. I. Muhammed, V. V. Priya, and V. A. Akhila, “A Novel Approach to Brain Biometric User Recognition,” Procedia Technol., vol. 25, no. Raerest, pp. 240–247, 2016. [18] P. Podder, M. M. Hasan, M. R. Islam, and M. Sayeed, “Design and Implementation of Butterworth, Chebyshev-I,” Int. J. Comput. Appl. (0975 – 8887), vol. 98, no. 7, pp. 12–18, 2014. [6] M. Phothisonothai, “An investigation of using SSVEP for EEGbased user authentication system,” in 2015 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2015, 2015, no. December, pp. 923–926. [19] Shakshi and R. Jaswal, “Brain Wave Classification and Feature Extraction of EEG Signal by Using FFT on Lab View,” Int. Res. J. Eng. Technol., vol. 3, no. 7, pp. 1208–1212, 2016. [20] [7] E. Fly, “A better fingerprint scanner for frequent flyers,” Fortune Features, 2013. O. Dressler, G. Schneider, G. Stockmanns, and E. F. Kochs, “Awareness and the EEG power spectrum: analysis of frequencies,” Br. J. Anaesth., vol. 93, no. 6, pp. 806–809, 2004. [8] C. K. Lee, C. C. Chang, A. Johar, O. Puwira, and B. Roshidah, “Fingerprint changes and verification failure among patients with hand dermatitis,” JAMA Dermatology, vol. 149, no. 3, pp. 294–299, 2013. [21] B. Rajak, M. Gupta, D. Bhatia, and A. Mukhgerjee, “Power Spectral Analysis of EEG as a Potential Marker in the Diagnosis of Spastic Cerebral Palsy Cases,” Int. J. Biomed. Eng. Sci., vol. 3, no. 3, pp. 23–29, 2016. [9] M. Dolezel, M. Drahansky, J. Urbanek, E. Brezinova, and T. Kim, Influence of Skin Diseases on Fingerprint Quality and Recognition. 2011. [22] S. A.Unde and R. Shriram, “PSD based Coherence Analysis of EEG Signals for Stroop Task,” Int. J. Comput. Appl., vol. 95, no. 16, pp. 1–5, 2014. [10] M. Drahansky, M. Dolezel, J. Urbanek, E. Brezinova, and T. H. Kim, “Influence of skin diseases on fingerprint recognition,” J. Biomed. Biotechnol., vol. 2012, 2012. [23] [11] A. e Z. Uquete, B. Quintela, and J. ao P. S. Cunha, “Biometric authentication using electroencephalograms : a practical study using visual evoked potentials,” Electron. E Telecomunicacoes, vol. 5, no. 2, pp. 185–194, 2010. B. L. Rajak, M. Gupta, D. Bhatia, A. Mukherjee, S. Paul, and T. K. Sinha, “Power Spectrum Density Analysis of EEG Signals in Spastic Cerebral Palsy Patients by Inducing r-TMS Therapy,” J. Biomed. Eng. Technol., vol. 4, no. 1, pp. 7–11, 2016. [24] P. K. Rahi and R. Mehra, “Analysis of Power Spectrum Estimation Using Welch Method for Various Window Techniques,” in International Journal of Emerging Technologies in Engineering (IJETE), 2014, pp. 106–109. [25] H. U. Amin, W. Mumtaz, A. R. Subhani, M. N. M. Saad, and A. S. Malik, “Classification of EEG Signals Based on Pattern Recognition Approach,” Front. Comput. Neurosci., vol. 11, pp. 1–12, 2017. [12] P. T. Dũng, Đ. H. Gia, L. Khải, and Đ. T. H. Vân, “EEG Signals For Authentication In Security Systems,” Nghiên cứu Khoa học và Công nghệ trong lĩnh vực An toàn thông tin, vol. 2, no. 03, pp. 17–32, 2016. [13] Z. Mu, J. Hu, and J. Min, “EEG-Based Person Authentication Using a Fuzzy Entropy-Related Approach with Two Electrodes,” Entropy, vol. 18, no. 12, 2016. [26] F. Pereira, T. Mitchell, and M. Botvinick, “Machine learning classifiers and fMRI: A tutorial overview,” Neuroimage, vol. 45, no. 1, pp. S199–S209, 2009. [14] E. Piciucco, E. Maiorana, O. Falzon, K. P. Camilleri, and P. Campisi, “Steady-State Visual Evoked Potentials for EEG-Based Biometric Identification,” in Lecture Notes in Informatics (LNI), Proceedings - Series of the Gesellschaft fur Informatik (GI), 2017. [27] M. Mustafa, M. N. Taib, Z. H. Murat, and N. Sulaiman, “Comparison between KNN and ANN Classification in Brain Balancing Application via Spectrogram Image,” J. Comput. Sci. Comput. Math., vol. 2, no. 4, pp. 17–22, 2012. [15] K. Mohanchandra, L. GM, P. Kambli, and V. Krishnamurthy,