Assessment of Human Random Number Generation for Biometric Verification ABSTRACT

Transcription

1 Original Article Assessent of Huan Rando Nuber Generation for Bioetric Verification Elha Jokar, Mohaad Mikaili Departent of Engineering, Shahed University, Tehran, Iran Subission: Accepted: ABSTRACT Rando nuber generation is one of the huan abilities. It is proven that the sequence of rando nubers generated by people do not follow full randoness criteria. These nubers produced by brain activity see to be copletely nonstationary. In this paper, we show that there is a distinction between the rando nubers generated by different people who provide the discriination capability, and can be used as a bioetric signature. We considered these nubers as a signal, and their coplexity for various tie-frequency sections was calculated. Then with a proper structure of a support vector achine, we classify the features. The error rate, obtained in this study, shows high discriination capabilities for this bioetric characteristic. Key words: Approxiate entropy, rando nuber generation, support vector achine, verification bioteleetry, wavelet decoposition Introduction The Longan Dictionary of Conteporary English describes rando as happening or chosen without any definite plan, ai, or pattern. A sequence of nubers is said to be rando, if the next eleent cannot be predicted fro the previous one. [1] For any years, scientists have been checking the ability of huan beings to generate true rando nubers. During the rando nuber generation, each subject has his/her own strategy and ust partially eorize the previous nubers and choose the next one based on his/her own conception of randoness. Since 1960, Alan B. Baddeley investigated this field extensively. Recently, new statistical processing ethods have been applied to discover this ability of huan beings and in ost of the an has been known as a bad rando nuber generator. [1-3] Generally, generation of rando nubers activate certain areas of the brain. If nubers are not written down, then the ind can only review those nubers held in short-ter eory which causes pattern suppression. [3] Because of this process, the rando sequence generated by people, is biased copared to true rando nubers. [1] Analysis of rando nubers generated by people to investigate short- ter eory function and attention is a well-explored area of psychological research. [3] When soeone is asked to generate rando nubers, a cognitive load is iplied, since there is a close interaction between executive eory and internalized decisionaking echaniss. Several studies in cognitive psychology show that the generation of rando rhyths is a cognitive task and has enough inforation for discriination of different clinical populations. A closely related task to nuber generation is rando rhyths of tapping a key. [4] The first assessent of this idea is done in a study by Hornero et al. [5] in classification of schizophrenic patients fro healthy subjects. He showed that the tapping rhyth of schizophrenic patients differ fro healthy subjects (P ANOVA < 0.001). The tie series generated by tapping a key for schizophrenic patients had a lower coplexity and variability than the healthy subjects. He used chaotic dynaic attractors and the secondorder difference plots of the tie series, and calculation of the correlation diension and the central tendency easure (CTM) paraeter. Then he concluded that this test could be a copleentary tool to help physicians in the estiation of cognitive-otor dysfunction in schizophrenic patients. Address for correspondence: Ms. Elha Jokar, Bioedical School, Shahed University, Persian Gulf Way, Tehran, Iran. E-ail: elhajokar@gail.co 82 Vol 2 Issue 2 Apr-Jun 2012

2 In 2006, Hornero et al. [6] analyzed the tie series generated by schizophrenic patients and control subjects using three nonlinear ethods of tie series: CTM fro the scatter plots of first differences of data, approxiate entropy (ApEn), and Lepel-Ziv (LZ) coplexity. He constructed a training set and a test set and used the training set for algorith developent and optiu threshold selection. Each ethod was assessed using the test dataset. He obtained 80% sensitivity and 90% specificity with LZ coplexity, 90% sensitivity, and 60% specificity with ApEn, and 70% sensitivity and 70% specificity with CTM. His results show differences in the ability to generate rando tie series between schizophrenic subjects and controls, as estiated by the CTM, ApEn, and LZ. Until this tie, rando rhyths just used to discriinate between clinical and control subjects, but in 2009, Laskaris et al. [4] has suggested a new ethod in bioetric verification with the repetitive pressing of a button in a rando anner. Bioetrics includes ethods for recognizing a person, based on a physiological characteristic, like fingerprints, face, hand geoetry, handwriting, iris, and voice. As the level of security violations and transaction fraudulent increases, the need for highly secure identification and personal verification technologies is becoing apparent. Bioetric solutions are able to provide for confidential financial transactions and personal data privacy. Utilizing bioetrics for personal identification is becoing considerably ore reliable and convenient than current ethods. Current ethods such as the utilization of passwords ay be used by soeone other than the authorized user and also it is required to eorize, but bioetrics links the event to a particular individual and because of this reason it has the low possibility of error and fraud. The first wave of bioetrics includes natural characters were distinctive and static, like recognizing a person, based on his\her fingerprints, iris, hand geoetry, vein topography. When the fraud caused the loss of security, the scientists tried to find the second wave of bioetrics with ore dynaic characteristics like the voice and handwriting style which are difficult to be iitated. One of the bioetric approaches in this kind is keystroke dynaics which is known as typing recognition. It analyzes the way a person types, since no extra hardware is required and typing is the ost natural way for a user to interact with the syste in ost applications, particularly over the world-wide web. However, the passage to be typed ight need to be fixed and this eans a eory load, siilar to reebering an extra password. Moreover, there is a potential change due to continuous practice of the sae typing patterns. [7] The new bioetric character introduced by laskaris was the rando tie intervals of tapping a key, in which the siplicity of interface is kept, while the restriction of typing specific patterns is alleviated. Key stroke dynaic is a cognitive task, since it depends on higher brain functions that can be indirectly easured in rando generation. Interestingly, he deonstrated that everyone has his own Eigen-rhyths regulating spontaneous finger tapping. Laskaris used this ability of huans to verifying the person s identity. His proposed ethod was based on graph theory and calculated siilarity between two signals for identification of their generators. Their study could achieve 93% accuracy with support vector achine (SVM), 95% accuracy with iniu class variance support vector achine (MCVSVM) classifiers. [4] In this paper a new ethod is suggested for distinction between the rando generations of individuals. This ethod has been used for chaotic quantification of cerebral signals. To deonstrate the effectiveness of this ethod, two different protocols were designed. The first protocol is based on laskaris experients. The second test, which is a new protocol in the field of bioetrics, was established on huan conception of the rando nubers. When soeone is asked to generate (verbally of via keyboard) rando nubers, there is a cognitive load iplied, since there is a close interaction between short-ter eory and internalized decision aking echaniss. This ability of huans is very valuable, especially in a world where counterfeiting and fraud has been rapant and the need for highly secure identification and personal verification technologies is becoing obvious. In this study, we test a new ethod for identification people fro their rando tapping and verify if a discriination scenario which proposed, is possible by rando generation of nubers. To answer these questions, two different experients were established: - Subjects should press the keyboard keys in a rando anner. The tie intervals between the hits are considered as a signal for that subject. - In the second test, we choose a special range of the nuber space (1-9) as our original set, and subjects were asked to keep this set in ind and generate rando nubers verbally. Due to the nonstationary characteristics of the signals in both experients, we used Coplexity assessent ethods to distinguish between nuber sets generated by individuals. Part 2 describes the conditions of the experients. In section 3 the ethod of feature extraction will be explained. Section 4 includes the results and final section contains a brief discussion on the results obtained in this study. PROTOCOLS AND PARTicipanTS A total of 30 subjects (6 ale and 24 feale aged 24.6±1.505) Vol 2 Issue 2 Apr-Jun

3 participated in this experient. The subjects did the tests while they were seated relaxed in a quite laboratory. During the tests, only one participant attended in the laboratory at a tie. This restriction was iposed to iniize the environental effects their perforance. Each participant was asked to do the tests 10 ties in five different days. In each session, two parts were perfored. Each part consists of two protocols: nuber test and keyboard test. Subjects had a short break for a few inutes between each two parts in one session. Nubers Test Due to the ental perception of subjects of decial nuber space, the selection range was set fro 1 to 9, and the subjects ust generate a sequential series of rando nubers verbally in a liited tie (140 seconds per test). The eaning of the rando nubers was explained with the theory of hat and nine nubered balls in it. The description of randoness was the sae for all subjects, and they were not allowed to ask any questions. To generate a rando nuber, the subject should iagine taking a ball out of the hat, read its nuber, and return it. In this study we do not use the external paced for nuber generation and subjects should generate nubers with their own rhyth. We considered tie liitation in this protocol because our processing was independent of the length of the signals. [Figure 1] shows two exaples of rando nuber signals recorded fro two different subjects. During the test, voice of the subjects was recorded and transcribed later by the experienter on a personal coputer and converted to a MATLAB signal. Keyboard Test In this test, subjects were asked to press the keys of the keyboard with the index finger of their doinant hand in an as rando anner as possible, until the screen shows the end of the exercise. In first session for each subject the eaning of rando tie interval for pressing the space key was explained with an exaple consist of a square 4 4 c, which appears and disappears in the screen at rando rhyth. In this study the nuber of strokes were T=150. This increase in length of the signals in coparison with previous work is to eliinate the effect of subjects fatigue for two consecutive protocols (nuber test and keyboard test). If X[n] denotes the sequence of exact tie-latencies of subjects hits X[n] = [t 1,t 2,...,t T ], the corresponding signal takes the for x[n] = [t 2 t 1,t 3 t 2,...,t T t T 1 ]. After each test, this series was reconstructed and used as input signal. [Figure 1b] shows two rando tie interval for two subjects. SIGNAL PROCESSING The processing was perfored in two stages: A. Feature extraction B. Classification. In Figure 2 a scheatic of our signal processing ethod is shown. In following sections each of these steps is described. Feature Extraction In the first stage, features are extracted fro the data using tie frequency doain ethods. Since the signal has nonstationary characteristics in general, it is ore appropriate to use tie frequency doain ethods like wavelet transfor for feature extraction. Wavelet transfor does not require the assuption of quasi-stationary on the data. It supports both tie and frequency aspects of a signal siultaneously, which akes it possible to capture accurately and localize transient characteristics of the data. Signals were decoposed into various frequency bands through wavelet packet decoposition. Then, the a b Figure 1: Exaple of signals generate by two subjects in different tests, (a) nuber test for two subjects with different length, (b) keyboard test for different subjects 84 Vol 2 Issue 2 Apr-Jun 2012

4 x = { y, y, y,..., y }, 1 N ( 1 ) (1) i i i + i + 2 i + ( 1) where N is the length of the nuber signal and τ and are the tie delay and ebedding diension, respectively. The distance between two points of the trajectory can be considered as the axiu difference in their corresponding eleents: dxi ((), x()) j = ax( yi ( + ( k 1) ) y( j+ ( k 1) ) ) k = 12,,.., Then, the siilarity between a point of the trajectory and the others are coputed as (2) Figure 2: Scheatic of our signal processing ethod C 1 () r = i N ( 1τ ) θ ( r d( r d( xi (), xj ( ))) (3) j i approxiate entropy (ApEn) -- subsequently described -- of the wavelet coefficients is calculated at the various nodes of the decoposition tree and used as a feature set for training the proper classifier. ApEn is a easure of predictability or regularity of a tie series and helps to understand the underlying chaotic behavior of the brain that shows itself in the nubers. Wavelet packet decoposition (WPD): Wavelet packet analysis is a generalized for of the discrete wavelet transfor. In the wavelet packet analysis, the signal is first passed through a low-pass and a high-pass filter, in parallel. The cut-off frequencies of these filters are onefourth of the sapling frequency. The bandwidth of the filters is half the bandwidth of the original signal, which allows downsapling of the output signals by two without losing any inforation according to the Nyquist theore. At each level of the decoposition, frequency resolution is doubled through filtering while the tie resolution is halved by downsapling operation. [8] In this study, signals were decoposed into various frequency bands through a two level wavelet packet decoposition. Based on the shape of signals, Daubechi2 (Db2) other wavelet is used for the decoposition. After this two level decoposition, we have seven vectors of wavelet coefficients for each signal; we then calculate ApEn for each of the vectors and construct the feature atrix. Approxiate entropy (ApEn): ApEn which is derived fro the Kologorov Sinai entropy is a tool for calculation of the coplexity of a signal. It is defied as the log likelihood (how likely is) that runs of patterns of certain length that are close to each other will reain close on next increental coparisons. [8] A deterinistic signal is expected to have a saller ApEn value than a highly irregular or rando one. To copute the ApEn of a signal, y i,i=1,,n, the trajectory in the ebedding space, R, ust first be reconstructed using the reconstruction ethod: where θ(x)=1 for x>0, and θ(x)=0 otherwise, and r is the vector coparison threshold. Finally, we define Φ (r) as ϕ N 1 () r = N ( ) τ i= ( 1) τ 1 1 log C ( r). i For fixed, r an τ, ApEn is defined as ApEn( r,, τ, N) = ϕ ( r ) ϕ +1 ( r) The ApEn is calculated for the wavelet coefficients obtained in the previous step and are used as our features. Thus, for each signal, seven vectors were obtained by WPD and Entropy was calculated for each of the. Vector coparison distance (r) and the tie delay (t) were set to 0.15 ties the standard deviation of the coefficients and 1, respectively. Different results were obtained using different values for. Classification After construction feature atrix fro ApEn, it can be used for training a suitable classifier. For each of the 30 participants, there were 10 signals, so totally we had 300 signals with seven features for each. Therefore, the size of the feature atrix was Due to the large nuber of classes (every subject is considered as a class), and a sall nuber of data for each class, neural network was not recognized as a suitable classifier. However, the support vector achine (SVM), has the required characteristics and was used as our classifier. The SVM ethod is based on the solution of a quadratic optiization proble that represents a trade-off between the iniization of the epirical error and the axiization of the soothness of the regression function. In the present work, different kernels were used for classification to choose the best one. To solve the proble of lack of data in each class and avoid overfitting the classifier, the standard criteria Leave-one-out (LOO) was (4) (5) Vol 2 Issue 2 Apr-Jun

5 used. In this way, for each subject, two signal left out to be used as a test set and the reaining, used for training. To coplete the test set and to have an ipostor clai, we oved over the 30 subjects and excluded all saples of one person fro the training set in each turn along with one session (two signals) fro each of the other 29 subjects. Therefore, the training set consists of all but one of the 30 subjects. The two signals fro each of the 29 subjects which was excluded fro the training set was used as our test set along with all the data of the iposter (the subject who was excluded fro the training set). Let A 1, A 2, A 3,..., A 30 be the identity codes of the subjects included in the database. Figure 3 depicts the experiental protocol, when A 1 is considered to be the Ipostor and session 5 is eployed as test set. It can be seen that the training set is built of four out of the five available sessions each one consisted of 29 out of the 30 available subjects. In the classifier, each subject considered as a separate class, and the SVM was built based on the concept of one against all. So, we build a classifier consist of twenty nine SVM structures each trained for recognition of one class against others. Results Thirty volunteers participated in this study; two tests were established to validate the proposed ethod. In each test, 10 signals captured fro each individual. For each signal seven features were extracted, after applying two-level wavelet decoposition and calculating the approxiate entropies, the classification scenario was trained and tested as described in previous section. In this case, the acceptance rate and the rejection rate were considered equal. The error rate was calculated as the average of the errors in all runs. The results for different kernels for nuber test, is given in Table 1. The paraeter is related to how to calculate the approxiate entropy and shows the ebedding diension in calculation phase. It is shown that, SVM achieved a very good result of 4.3% average error rate in perforance verification using linear and RBF kernels with siga=5. Results of keyboard test and nuber test are copared in Figure 4. As can be seen, a better accuracy is achieved in the nuber test copared to the keyboard test. However, both tests got the best results with Siga=5. Conclusion In the past, the rando nubers generated by huans were only used to activate certain parts of the brain. This study showed that behind the irregular and chaotic behavior of nubers produced by people, there is a distinctive feature which is very interesting and useful and can be used to identify the producer. This ability was seen in the tie series obtained by pressing a key in clinical populations. In those studies distinction between healthy and patient populations were perfored using signals fro rando tie intervals of tapping a key. In 2009, Laskaris showed that the interval between pressing the key has the features that vary between different individuals. In laskaris experients, adding diensions of tie series was reconstructed and the atrix of siilarity between the signals using the iniu spanning tree and the ultivariate Wald- -Wolfowitz test was calculated. In his work, with equal acceptance rate and rejection rate he could achieve 7.55% error with SVM, and 5.44% error with MCVSVM structure. Coparison of responses obtained in this experient and Laskaris test are suarized in Table 2. The results Figure 3: Verification scenario Table 1: Error rate of verification with chenge of kernels Kernel function =2 Error (%) =3 Error (%) Rbf Siga= Siga= Siga= Siga= Figure 4: Results of SVM based scenario for keyboard test and nuber test with RBF kernel Table 2: Coparison between errors obtained until now Bioetry protocol Laskaris test Feature extraction ethod Ebed diension+ MST + WW-test + MDS The best classifier Error (%) MSVSVM 5.40 Keyboard test Wevelet Dec. +ApEn SVM 4.7 Nuber test Wevelet Dec. +ApEn SVM 4.3 Vol 2 Issue 2 Apr-Jun 2012

6 are rearkably siilar, although the ethod used in this article and laskaris ethod is copletely different. This has two iportant aspects: Firstly, the ethod used in this study gave the correct answer to signals with a protocol siilar to Laskarys protocol (using keyboard pressing instead of tapping one key). As a result our ethod has been properly established and ipleented. Secondly, the ability of generating rando nubers can be confired as a bioetric feature. As can be seen the rando nubers ake better distinction between their producers copared to rando rhyths. This higher accuracy ay be due to nubers range fro 1 to 9 that causes ore clear perception of randoness in people and then specified ental processing on the data. Despite of lower error rate in nuber test, this protocol is easier to fraud than the keyboard test. Then we can say that keyboard test is safer than the nuber test. One of the technical challenges that need to be considered is converting of the speech signal to nuber signal in nuber test. The keyboard test does not have this part. With all probles listed above, results in this paper show nuber test as a keyboard test can considered as bioetry character of each person. A significant practical issue is the adequate nuber of signals in each protocol. In previous experients the lengths of key tapping were 128 hits; however only one test was taken fro the participants. In this study, subjects first participated in the nuber test and then did the keyboard test. So to eliinate the effects of exhaustion fro the previous test, and also to ake a clear interpretation of the brain processes, the signal length was increased. This can be one of the reasons for error reduction in this protocol. In the nuber test due to liitation in test tie the nubers were produced by individuals were different. This variation did not affect the ipleentation process. However soe of the individuals are very slow in nuber generation and their signals are short. Then if the length of data for all individuals were got equal in this protocol, perhaps a greater accuracy could be expected. This bioetric feature is very valuable, because it requires no special tool and its process is so quick and easy. Especially, the possibility for fraud and fake on it is very difficult. In the world of the Internet and webs where people should be identified virtually and reotely, it would be very useful. The nuber of participants in this experient does not allow us to conclude in general, but as a preliinary study it is very satisfactory. Of course, if this field of bioetrics can reach an acceptable accuracy, this protocol could becoe a coprehensive tool in security identification systes. References 1. Figurska M, Czyk MS, Kulesza K. Huans cannot consciously generate rando nubers sequences: Poleic study. Med Hypotheses 2008;70: Persaud N. Huans can consciously generate rando nuber sequences: A possible test for artificial intelligence. Med Hypotheses 2005;65: Bains W. Rando nuber generation and creativity. Med Hypotheses 2008;70; Laskaris NA, Zafeiriou SP, Garefa L. Use of rando tie intervals (RTIs) generation for bioetric verification. Pattern Recognit 2009;42: Hornero R, Alonso A, Jieno N, Jieno A, Lopez M. Nonlinear analysis of tie series generated by schizophrenic patients. IEEE Eng Med Biol 1999;18: Hornero R, Abà solo D, Jieno N, Sà nchez CI, Poza J, Aboy M. Variability, Regularity, and Coplexity of Tie Series Generated by Schizophrenic Patients and Control Subjects. IEEE Trans Bioed Eng 2006; Vol Peacock, Xian K, Wilkerson M. Typing patterns: a key to user identification. IEEE Secur Priv 2004;2: Ocak H. Optial classification of epileptic seizures in EEG using wavelet analysis and genetic algorith. Signal Process 2008;88: How to cite this article: *** Source of Support: Nil, Conflict of Interest: None declared Biographies Elha Jokar, MSc student in Bioedical Engineering, Shahed University, Tehran, Iran, E-ail: elhajokar@gail.co Mohaad Mikaili, Bioedical Professor, Shahed University, Tehran, Iran E-ail: _ikaili@yahoo.co Vol 2 Issue 2 Apr-Jun