2.1 Introduction to person identification systems
There are a lot of systems require reliable personal recognition schemes to either con-firm or determine the identity of an individual requesting their services. The purpose of such schemes is to ensure that the supplied services only can be accessed by a allowed user and no one else. Examples of such applications include secure access to buildings, computer systems, laptops, cellular phones and ATMs [12].
Conventional person identification methods include passwords, smart cards, and a vari-ety of biometric techniques. Passwords and smart cards are widely-used because of the ad-vantage of convenience. However, smart cards might be stolen, simple passwords might be deciphered, and complicated passwords might be forgotten. Biometric recognition, refers to the automatic recognition of individuals based on their physiological or behavioral char-acteristics is popular recently and considered more secure way.
Biometric recognition
By using biometrics it is to confirm or establish an individual’s identity based on ”who he/she is”, rather than by ”what he/she possesses” (e.g., a smart card) or ”what she remem-bers” (e.g., a password). What qualities need to have a biometric can be used in identity?
Any human physiological or behavioral characteristic can be used as a biometric character-istic on condition that it satisfies the following requirements:
• Universality: anyone have this characteristic;
• Distinctiveness: the characteristics of any two people should have sufficient differ-ences to separate different people;
• Permanence: the characteristic should be sufficiently invariant with time (correspond to the matching criterion);
• Collectability: the characteristic can be measured quantitatively.
In addition to the above requirements, a practical biometric system should be provided with sufficient accuracy and speed, be user-friendly, and it can expand the number of users. It is also necessary to prevent the impostors by robust anti-theft mechanism.
2.2 Categories of biometrics 13
Figure 2.1: Examples of biometric characteristics. From left to right are fingerprint, iris, voice, palm and gait.
Biometric system
A biometric system is essentially a pattern recognition system that works by biomet-ric data acquired from an individual. After feature extracting from the acquired data the features are compared to the template set in the database.
The biometric system can be divided into two mode depending on the application. One is verification mode that the system validates a person’s identity by comparing the acquired biometric data with his/her own biometric template stored in database. People who want to be approved by the system need to claim an identity via a PIN (Personal Identification Number), a user name or a smart card. The system conducts a one-to- one comparison to determine whether the claim is true or not (e.g., user: I am John. system: Whether this biometric data belong to John.). The other is recognition mode that the system recognizes an individual by searching the templates of all the users in the database for a match. A one-to-many comparison to establish an individuals identity consequently conducted in this system with out a declaration of subjects; however, it will fail if the subject is not enrolled in the database (e.g., system: Whose biometric data is this?).
2.2 Categories of biometrics
Fingerprint
In 19thcentury, Alphonse Bertillon who is a chief of the criminal identification division conceived and then practiced the idea of using a number of body measurements to identify criminals [23]. After this, a more significant and practical discovery of the distinctiveness of the human fingerprints became clear and soon the fingerprints used for criminal
iden-tification. While law enforcement agencies were the earliest adopters of the fingerprint identification technology, it is being increasingly used later because more identity fraud has created.
Iris
According to human eye micro vascular, infinite variety of combinations, the complex iris texture carries very distinctive information useful for personal recognition. The combi-nation of iris blood vessels will not change basically even with the age increasing (except for severe diabetes, glaucoma). It is difficult to surgically tamper because of the living conditions (such as vascular blood flow). In view of those characteristics of iris it become popular in person recognition.
Voice
Unlike fingerprint and iris, the biometric of voice is using signal characteristics. Voice is a combination of physiological and behavioral biometrics. The features of an individual voice are based on the shape, amplitude and frequency. The physiological characteristics of human speech are invariant for an individual. Speaker recognition is most appropriate in phone-based applications such as voice dialling. However, there are some disadvan-tages lead to difficulties in the application of person identification (we will detail in next subsection).
2.2.1 Disadvantages of present-day biometric systems
There are other biometrics such as face, palm and gait except above described. Al-though biometrics are more reliable and secure method of identity, it still has some dis-advantages. For fingerprint, people long-term need to work by hands might lose their fin-gerprint (e.g., repair worker). For iris recognition, it is necessary to use infra-red scanning eye that may cause safety concerns. Voice is not very distinctive and it can be imitated by training. The behavioral part of the speech of a person might change over time, and is also influenced by emotion and different physical states (such as common cold). Furthermore,
2.3 EEG-based person identification systems 15
collection of voice depend on the quality of microphone and is sensitive to background noise.
In addition to the disadvantages above mentioned of different biometrics, present-day biometrics can be stolen, duplicated, or even provided under violent threats. Therefore, a more robust person identification system is necessary.
2.3 EEG-based person identification systems
To address the disadvantages of existing biometrics some researchers brought up using brain signal as a biometric [21, 22]. To evaluate the uniqueness and consistency of the characteristics in EEG signal, the work in [18] 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 intra-subject were repeatedly acquired for several times. The characteristics in EEG signal fit well with the requirements of biometric system.
2.3.1 Basic components of EEG-based person identification
Signal pre-processing
Typical procedures include amplification, filtering and artificial rejection in order to improve the signal-to-noise ratio. In general, the brain activity is obscure and difficult to detect, a sufficient amplification consequently required. The bandpass filter is usually applied for filtering to cover high pass and low pas filter. In addition, a notch filter is also used to suppress the 60 Hz power line interference. For artificial rejection, the electro-oculographic (EOG) and electromyographic (EMG) are excluded detected by a predened threshold.
Feature extraction
The original signal of brainwave is chaotic that lead to difficulties in the practical ap-plication. In addition to common amplitude and latency information are used in analysis
various feature extraction methods have been studied to extract more discriminative fea-tures, such as discrete wavelet transform, continuous wavelet transform, autoregression model (AR), power spectrum and so on.
Classification
The results of classification is to perform how suitable the feature used for distinguish-ing between different people. Many classification methods have been proposed in pattern recognition field. The classifier in a person identification can be anything from a simple linear model to a complex non-linear or a machine learning models. The acquired data are divided into training phase and testing phase. The training phase consists of a repetitive process of tasks to train a classifier and then use the testing phase to evaluate the perfor-mance.
2.3.2 EEG signals: resting data
The acquired signal of brainwave are mainly divided into two types, resting data and task dependent data. The resting data has the advantages of easy operation and when people in resting state the brain will generate the alpha rhythm that can be used as a waveform characteristic of each subject. In 2002, Poulos proposed a bilinear model to find the non-linear components in the EEG and the identification rate ranged from 72 to 85% [20]. By a lot of methods of feature extraction such as autoregressive (AR) coefficients, coherence and cross-correlation, the performance analysis of the system that Riera proposed obtained true acceptance rate of 96.6% [24]. For simplicity and practicability, the work [16] classified subjects simply by thresholding the EEG power spectrum.
2.3.3 EEG signals: task dependent data
Compared to resting data, the task-induced EEG signals are more specific that will re-duce spontaneous effects of signal analysis. In 2003, Palaniappan and Ravi investigated the task-related EEG signals [19], the stimuli used in their work were standard image database [27]. By extracting features(channel wise power spectral density) from visual
2.4 Limitations of biometric systems 17
Figure 2.2: Task dependent experiment. To induce EEG of task dependent data there are three examples: From left to right are standard image database, checkerboard and motor imagery.
evoked potentials (VEPs), the identification accuracy was improved to be larger than 93%.
A novel peak matching algorithm proposed by Singhal that only relied on recording from single channel gave 78% accuracy [26], they used a checkerboard pattern to induce steady-state visually evoked potentials (SSVEP). Moreover, Marcel [14] devised more appropriate mental tasks which contained imagination of repetitive self-paced hand movements and generation of words to perform their research. A statistical framework based on Gaussian mixture models and maximum a posteriori model adaptation successfully applied to person authentication.
2.4 Limitations of biometric systems
Noise
A fingerprint with a scar, or a voice altered by cold are examples of noisy data. Simi-larly, the brainwave might be affected by emotions; however, this disadvantage may be an advantage for person identification that prevent impostors from threats. The old measuring instruments or unfavorable ambient conditions such as poor illumination in face recognition system will reduce the accuracy of discrimination.
Intra-class variations
Although the biometric is considered stationary, it may be very different from the data used to generate the template. This variation is typically caused by a user who is incor-rectly interacting with the sensor. For instance, the different angle used in face recognition.
Sensor replacement might cause different results of the same subject. The varying psycho-logical makeup of an individual might result in different behavioral characteristic at various time measurements.
Moral issues
While the biometric is a more reliable and secure method of identity, some people may be reluctant to provide part of their body to as a biometric that they probably feel that privacy has been violated. It is important that biometrics acquiring needs to solicit for user’s consent. To a large extent, the human factor dominates the success of the biometric-based identification system. Therefore, the biometric system must be user-friendly and accepted by users.
2.5 Thesis scope
After reviewing related researches for person identification, we attempt to adopt the method of training the EEG of subjects as a biometric to identify different persons. We designed a simple task that formed by different-sized disc and presented in different pro-portions of occurrences to induce the VEPs and ERPs. The feature extraction of raw EEG data which correspond with the classifier and can achieve the better classification is the major part of our work. The proposed EEG-based person identification system will test by 18 subjects to verify its reliability. The methods of feature extraction described in the Chapter 3, we used the features to train a accurate classifier which contained classification phase and verification phase.
The preliminary classification obtained by a multi-class classifier. The best-matching candidate of each classification is further verified by using a binary classifier to exclude the impostors in verification phase. For the performance evaluation, the accuracy rate and the error rate are were used. Besides, we tried to correct the false classified data by the confidence value of the SVM classifier. Summarizing the methods described above, we developed a reliable and accurate EEG-based person identification system.