Unobtrusive Biometric System Based on Electroencephalogram Analysis

Features extracted from electroencephalogram (EEG) recordings have proved to be unique enough between subjects for biometric applications. We show here that biometry based on these recordings o ﬀ ers a novel way to robustly authenticate or identify subjects. In this paper, we present a rapid and unobtrusive authentication method that only uses 2 frontal electrodes referenced to another one placed at the ear lobe. Moreover, the system makes use of a multistage fusion architecture, which demonstrates to improve the system performance. The performance analysis of the system presented in this paper stems from an experiment with 51 subjects and 36 intruders, where an equal error rate (EER) of 3.4% is obtained, that is, true acceptance rate (TAR) of 96.6% and a false acceptance rate (FAR) of 3.4%. The obtained performance measures improve the results of similar systems presented in earlier work.


INTRODUCTION
The term "biometrics" can be defined as the emerging field of technology devoted to identification of individuals using biological traits, such as those based on retinal or iris scanning, fingerprints, or face recognition.
Biometrics is nowadays a big research playground, because a highly reliable biometric system results extremely interesting to all facilities where a minimum of security access is required. Identity fraud nowadays is one of the more common criminal activities and is associated with large costs and serious security issues. Several approaches have been applied in order to prevent these problems.
New types of biometrics, such as EEG and ECG, are based on physiological signals, rather than more traditional biological traits. This has its own advantages as we will see in the following paragraph. An ideal biometric system should present the following characteristics: 100% reliability, user friendliness, fast operation, and low cost. The perfect biometric trait should have the following characteristics: very low intrasubject variabil-ity, very high intersubject variability, very high stability over time and universal. Typical biometric traits, such as fingerprint, voice, and retina, are not universal, and can be subject to physical damage (dry skin, scars, loss of voice, etc.). In fact, it is estimated that 2-3% of the population is missing the feature that is required for the authentication, or that the provided biometric sample is of poor quality. Furthermore, these systems are subject to attacks such as presenting a registered deceased person, dismembered body part or introduction of fake biometric samples.
Since every living and functional person has a recordable EEG signal, the EEG feature is universal. Moreover, brain damage is something that rarely occurs. Finally, it is very hard to fake an EEG signature or to attack an EEG biometric system.
The EEG is the electrical signal generated by the brain and recorded in the scalp of the subject. These signals are spontaneous because there are always currents in the scalp of living subjects. In other words, the brain is never at rest. Because everybody has different brain configurations (it is estimated that a human brain contains 10 11 neurons and 2 EURASIP Journal on Advances in Signal Processing 10 15 synapses), spontaneous EEG between subjects should be different; therefore a high intersubject variability is expected [11].
As it will be demonstrated with the results of our research, EEG presents a low intrasubject variability in the recording conditions that we defined: during one minute the subject should be relax and with his eyes closed. Furthermore, the system presented herein attains the improvement of the classification performance by combining a feature fusion with a classification fusion strategy. This kind of multistage fusion architecture has been presented in [22] as an advancement for biometry systems.
This paper describes a ready-to-use authentication biometric system based on EEG. This constitutes the first difference with already presented works [4,5,[7][8][9]. The system presented herein undertakes subject authentication, whereas a biometric identification has been the target of those works. Moreover, they present some results on the employment of EEG as person identification cue [4,5,[7][8][9], what herein becomes a stand-alone system.
A reduced number of electrodes have been already used in past works [4,5,[7][8][9] in order to improve the system unobtrusiveness. This fact has been mimed in our system. There is however a differential trait. The two forehead electrodes are used in our system, while in other papers other electrodes configurations are used, for example, [5] uses electrode P4. Our long-term goal is the integration of the biometric system with the ENOBIO wireless sensory unit [23,24]. ENOBIO uses dry electrodes, avoiding the usage of conductive gel and therefore improving the user friendliness. For achieving this goal employing electrodes in no hair areas becomes mandatory, a condition our system fulfils.
Lastly, performance evaluation is worth mentioning. Although we present an authentication system, we have conducted some identification experiments for the sake of comparison with already presented works [4,5,[7][8][9]. The system presented herein shows a better performance by a larger number of test subjects. This question is further analyzed.
In the following sections, the used authentication methodology will be presented. Section 2 presents the EEG recording protocol and the data preprocessing. Section 3 deals with the features extracted from the EEG signal. Section 4 describes the authentication methodology, Section 5 the results; and finally conclusions are drawn in Section 6.

EEG RECORDING AND PREPROCESSING
For this study, an EEG database recorded at FORENAP, France, has been used. The database is composed of recordings of 51 subjects with 4 takes recorded on different days, and 36 subjects with only one take. All subjects were healthy adults between 20 and 45 years. The delay between the 1st and the 4th recording is 34 ± 74 days, whereby the mediumterm stability of the system will be tested. The recording conditions were the same for all subjects: they were seated on an armchair in a dark room, with closed eyes and were asked neither to talk nor to move, and to relax. The recording duration was between 2 and 4 minutes. Only the 2 forehead electrodes (FP1 and FP2) were used for authentication; and an additional electrode that was placed in the left ear lobe was used as reference. The decision of using the frontal electrodes is due to projective integration with the ENOBIO system, which was presented in the former section. Indeed, the forehead is the most comfortable place where EEG can be measured.
The sampling rate for data acquisition was 256 Hz. A second-order pass band filter with cut frequencies 0.5 and 70 Hz was applied as the first preprocessing stage. A narrow notch filter at 50 Hz was additionally applied.
Once the filters were applied, the whole signal was cut in 4-second epochs. Artefacts were kept, in order to ensure that only one minute of EEG data will be used for testing the system.

FEATURES EXTRACTION
Among a large initial set of features (Higuchi fractal dimension, entropy, skewness, kurtosis, standard deviation, etc.), the five ones that show a higher discriminative power in the conducted preliminary works were used. These five different features were extracted from each 4-second epoch. These feature vectors are the ones that we will input in our classifiers.
We can distinguish between two major types of features: those extracted from a single channel (single channel features) and those that relate two different channels (the synchronicity features).
Mutual information (MI), coherence (CO), and crosscorrelation (CC) are examples of two-channel features related to synchronicity [19][20][21]. They represent some joined characteristic of the two channels involved in the computation. This type of features is used for the first time in an EEG biometry system.
All the mentioned features are simultaneously computed in the biometry system presented herein. This is what we denote as the multifeature set. This set will be fused in subsequent stages of the system. The features are described in more detail in the following subsections.

Autoregression
The EEG signal for each channel is assumed to be the output of an autoregressive system driven by white noise. We use the Yule-Walker method, also known as the autocorrelation method, to fit a pth-order AR model to the windowed input signal, X(t), by minimizing the forward prediction error in a least-square sense. This formulation leads to the Yule-Walker equations, which are solved by the Levinson-Durbin recursion. The AR model is represented by A. Riera et al.

3
In this model, the time series are estimated by a linear difference equation in the time domain, where a current sample of the signal X(t) is a linear function of p previous samples plus an independent and identically distributed (i.i.d) white noise input e(t). The average variance estimate of e(t) is 0.75 computed for all the subjects. a(i) are the autoregression coefficients. Preliminary results have shown the convenience of using an AR model with order 100.

Fourier transform
The well-known discrete Fourier transform (DFT), with expression where is the Nth root of unity, is used herein to compute the DFT of each epoch. In our case, N is equal to 1024 (256 Hz * 4 seconds). We retain thence the frequency band from 1 to 40 Hz so that all EEG bands of interest are included: delta, theta, alpha, beta, and gamma.

Mutual information
In probability theory and information theory, the mutual information (MI), also known as transinformation [12,21], of two random variables, is a quantity that measures the mutual dependence of the two variables. The most common unit of measurement of MI is the bit, when logarithms of base 2 are used in its computation. We tried different numbers of bits for coding the signal, choosing 4 as the optimal value for our classification purposes. The MI has been defined as the difference between the sum of the entropies within two channels' time series and their mutual entropy.

Coherence
The purpose of the coherence measure is to uncover the correlation between two time series at different frequencies [19,20]. The magnitude of the squared coherence estimate, which is a frequency function with values ranging from 0 to 1, quantizes how well x corresponds to y at each frequency.
The coherence Cxy(f ) is a function of the power spectral density (Pxx and Pyy) of x and y and the cross-power spectral density (Pxy) of x and y, as defined in the following expression: In this case, the feature is represented by the set of points of the coherence function.

Cross-correlation
The well-known cross-correlation (CC) is a measure of the similarity of two signals, commonly used to find occurrences of a known signal in an unknown one. It is a function of the relative delay between the signals; it is sometimes called the sliding dot product, and has applications in pattern recognition and cryptanalysis.
We calculate three CCs for the two input signals: The correlation ρXY between two random variables x and y with expected values μ X and μ Y and standard deviations σ X and σ Y is defined as where (i) E() is the expectation operator, (ii) cov() is the covariance operator.
In this case, the features are represented by each point of the three calculated cross-correlations. This feature is referred to as CC in the following section.

AUTHENTICATION METHODOLOGY
The work presented herein is based on the classical Fisher's discriminant analysis (DA). DA seeks a number of projection directions that are efficient for discrimination, that is, separation in classes.
It is an exploratory method of data evaluation performed as a two-stage process. First the total variance/covariance matrix for all variables, and the intraclass variance/covariance matrix are taken into account in the procedure. A projection matrix is computed that minimizes the variance within classes while maximizing the variance between these classes. Formally, we seek to maximize the following expression: where (i) W is the projection matrix, (ii) S B is between-classes scatter matrix, (iii) S W is within-class scatter matrix.
For an n-class problem, the DA involves n − 1 discriminant functions (DFs). Thus a projection from a ddimensional space, where d is the length of the feature vector to be classified, into an (n − 1)-dimensional space, where d ≥ n, is achieved. In our algorithm, we work with 4 different DFs: (i) linear: fits a multivariate normal density to each group, with a pooled estimate of the covariance; (ii) diagonal linear: same as "linear," except that the covariance matrices are assumed to be diagonal; 4 EURASIP Journal on Advances in Signal Processing (iii) quadratic: fits a multivariate normal density with covariance estimates stratified by group; (iv) diagonal quadratic: same as "quadratic," except that the covariance matrices are assumed to be diagonal.
The interested reader can find more information about DA in [13].
Taking into account the 4 DFs, the 2 channels, the 2 single channel features, and 3 synchronicity features, we have a total of 28 different classifiers. Here, we mean by classifier, each of the 28 possible combinations of feature, DF, and channel.
We use an approach that we denote as "personal classifier," which is explained herein, for the identity authentication case: the 5 best classifiers, that is, the ones with more discriminative power, are used for each subject. When a test subject claims to be, for example, subject 1, the 5 best classifiers for subject 1 are used to do the classification. In order to select the 5 best classifiers for the 51 subjects with 4 EEG takes, we proceed as follows. We use the 3 firsts takes of the 51 subjects for training each classifier, and the 4th take of a given subject is used for testing it. We repeat this process making all possible combinations (using one take for testing and the others for training). Each time we do this process, we obtain a classification rate (CR): number of feature vectors correctly classified over the total number of feature vectors. The total number of feature vectors is around 45, depending on the duration of the take. Once this process is repeated for all 28 classifiers, we compute a score measure on them, which can be defined as score = average(CR) standard deviation(CR) .
The 5 classifiers with higher scores out of the 28 possible classifiers are the selected ones. We repeat this process for the 51 subjects.
Once we have the 5 best classifiers for all 51 subjects, we can then implement and test our final application. We now proceed in a similar way, but we only use in each test the first or the second minute of a given take, that is, we input in each one of the 5 best classifiers 15 feature vectors. Each classifier outputs a posterior matrix (Table 1). In order to fuse the results of the 5 classifiers, we vertically concatenate the 5 obtained posterior matrices and take the column average. The resulting vector is the one we will use to take the authentication decision (in fact it is a probability density function (PDF); see Figures 1(a) and 1(b), where the 1st element is the probability that the single minute test data comes from subject 1 and the 2nd element is the probability that the single minute test data comes from subject 2, and so forth.
The last step in our algorithm takes into consideration a decision rule over the averaged PDF. We use two different thresholds. The first one is applied on the probability of the claimed subject. The second threshold is applied on the signal-to-noise ratio (SNR) of the PDF, which we define as where P(x i / x i ∈ C i ) is the probability that the single minute test data comes from.

RESULTS
In the first part of this section, we provide the results for our authentication system. Then, for the sake of comparison with related works, which only deal with identification, we also provide the results of a simplified version of the "personal classifier" approach. This approach works as an identification system, that is, the claimed identity of the user is not taken into consideration as an input.

Authentication system results
Three different tests have been undertaken on our EEGbased biometric system in order to evaluate its classification performance: (i) legal test: a subject belonging to thedatabase claims his real identity, (ii) impostor test: a subject belonging to thedatabase claims the identity of another subject belonging to the database, (iii) intruder test: a subject who does not belong to the database claims the identity of a subject belonging to the database.
We have used the data of the 51 subjects with 4 takes in the database for the legal and the impostor tests. For the intruder test, the 36 subjects with 1 take have been applied to the system. An easy way to visually represent the system performance is the classification matrices (Figures 2(a)  and 2(b)). These are defined by entries c i j , which denote the number of test feature vectors from subject i classified as subject j.
Taking into account that we have 4 test takes, and that we use both the first and the second minutes for testing, we have 4 * 2 * 51 = 408 legal situation trials (N leg ). In the case of the impostor situation, we have also 4 takes, we also use the first and the second minutes of each take, we have 51 impostors that are claimed to be the other 50 subjects from the database. Therefore, we have 4 * 2 * 51 * 50 = 20,400 impostor situation trials (N imp ). For the intruder situation, we have 1 test take from which we only use the first minute, so we have 1 * 1 * 36 * 51 = 1,836 intruder situation trials (N int ). We use the true acceptance rate (TAR) and the false acceptance rate (FAR) as performance measures of our system. They are defined for each individual subject in each trial situation as following: where c i j denote the classification matrix entries as defined in the previous section, N the number of subjects for each trial situation, either legal/impostor (N = 51) or intruders (N = 36). It is worth mentioning that for this second case, no TAR i can be defined. Table 1: Posterior matrix of the 15 FT feature vectors extracted from one minute EEG recording of subject 1. Each row represents the probabilities assigned to each class for each feature vector. We see that the subject is well classified as being subject 1 (refer to the last row). Notice that this posterior matrix represents a 9-class problem and our work is done for a 51 class problem.  Figure 1: PDF for normal situation for subject 10 (a) and for intruder situation (b). In (a), notice that if a probability threshold is set to 0.15, subject 10 will be authenticate only if he claims to be subject 10. In (a), the intruder would not be authenticated in any case.
The general system TAR is computed as the average over all subjects: The general FAR can be computed in an analogous manner for the two different groups of impostors (N = 51) and intruders (N= 36).
As it can be observed, we get two different FAR measures for the impostor and the intruder cases. These two measures are weighted averaged in order to obtain a unique FAR measure as follows: where FAR imp is the average of FAR i over the 51 impostors, FAR int is the average of FAR i over the 36 intruder We finally obtain an equal error rate (EER) measure that equals 3.4%. This value is achieved for a probability threshold equal to 0.02 and an SNR threshold equal to 2.36. In Figure 3, we can see the behavior of TAR and FAR for different SNR thresholds (with probablitiy thresholds fixed to 0.02).
Depending on the security level, different thresholds can be applied in order to make the system more inaccessible for intruders, but this would also increase the number of legal subjects that are not authenticated as shown in Figure 3.

Comparison in an identification task
It is easy to slightly modify the described system to work in an identification mode. Indeed, this "identification mode" is a simplification of the authentication one. Rather than using personalized classifiers for each subject, what we do now is to use the same 16 classifiers for all the subjects. Those classifiers are the ones that have more discriminative power among all subjects. They are given in the Table 2.
It is worth pointing out that a trivial classifier would yield a CR equal to 0.0196 (i.e., 1/number of classes, which in our case is 51). Moreover, the results obtained after fusing the different classifiers significantly improve the performance of the identification system as depicted in Figure 4. This improvement of performance is also achieved in the "authentication mode." Figure 4 shows the behavior of the TAR and FAR for our system in "identification mode." We can see that 3 different operating points are marked. Those are the values we will use for the comparison. Table 2 shows several results from other works along with the results of our current work, in 3 different operating points.

DISCUSSION AND CONCLUSIONS
An authentication biometric system based on EEG, using 2 frontal electrodes plus 1 reference placed at the left ear lobe, is described in this paper. The tested subject has to sit, close her eyes, and relax during one minute of EEG recording. The only inputs to the system are the one-minute EEG recording and the claimed identity of the subject. The output is a binary decision: authenticated or not. This authentication system   demonstrates to outperform the same system in "identification mode" (EER = 3.4% versus EER = 5.5%). The "identification mode" is adopted only to compare with precedent studies [4, 5, 7-9], since they deal only with identification.
The results of our system in "identification mode" outperform precedent works even though a larger database has been used to test our system. Intruders have also been used to test the intruder detection. We consider that the more innovative point in this study is the use of several features and the way they are personalized and fused for each subject. We focus on extracting the maximum possible information from the test takes, taking care of the unobtrusiveness of the system: with only one minute of recording, using only the two forehead channels, we obtain 28 different classifiers, from which the 5 ones with more discriminative power for each subject are selected. In order to have an even more reliable system, a multimodal approach would probably increase the performance considerably. We are investigating the possibility of applying an electrocardiogram (ECG)-based biometry simultaneously to the EEG [14][15][16][17][18]. Combining EEG and ECG biometric modalities seems to be very promising and will be discussed in a follow-up paper.
Another possible application that we are researching is whether the emotional state (stress, sleepiness, alcohol, or drug intake) can be extracted from EEG and ECG. In this case, besides the authentication of the subject, we could undertake his initial state validation. This would be a very interesting application for workers of critical or dangerous environments.
Finally, the usage of less than one minute of EEG data recording is being studied in order to make the system less obtrusive. This condition will be improved as well with the ENOBIO sensory integration.