AU2013100576A4 - Human Identification with Electroencephalogram (EEG) for the Future Network Security - Google Patents

Abstract

Abstract Human identification becomes huge demand for various applications in particular for the security related areas, such as identification for a network security. Electroencephalogram (EEG) signals are confidential and hard to imitate, since EEG signals are a reflection of individual-dependent inner mental tasks. Generally speaking, EEG signal has several advantages, such as (i) it is confidential as it corresponds to a mental task, (ii) it is very difficult to mimic and (iii) it is almost impossible to steal as the brain activity is sensitive to the stress and the mood of the person. Hence, an aggressor cannot force the person to reproduce his/her mental pass-phrase. In this patent we proposed a novel algorithm to create a spatial pattern of EEG signals obtained from the open public database such that the proposed algorithm can be used to identify people by EEG signals. In our EEG signal processing, we have analyzed 64-electrode EEG samples for two databases, one is for 45 people and calculate the equivalent root mean square (rms) values for each electrode signal over 1 second period, by which created a 64-value input for each subject. With this neural network (NN) model, our analysis clearly showed that our designed classifier is able to identify all the 45 people correctly (successful rate of 100%) with a mean square error of 2.0334x 10-7 and the same algorithm applying to the 2 "d database with 116 out of 122 people can be fully identified (successful rate of 95.1%) with a mean square error value of 0.00186. We deeply believe that a low complexity, high resolution, effective and efficient is very attractive for the real life applications especially for network security in the foreseeable future. Keywords: biometric nature, security system, neural network, EEG, signal processing Figure 1: 64 vectors transformed from corresponding 64 electrodes Final Mean Square Error =0.0027151 - Train Test Best C> 102 10 - L LI U 200 400 600 800 1000 1200 Epochs Figure 2: Mean square errors for the second database, 122 people (subjects) during the training

Description

1 DESCRIPTION OF THE ART [1] EEG signals are the signatures of neural activities. They are captured by multiple-electrode EEG machines either from inside the brain, over the cortex under the skull, or certain locations over the scalp, and can be recorded in different formats. [2] Up to the present, EEG signals have been successfully applied to the research and development of brain-computer interfaces whose main goal is to enhance the communication and control abilities of motor-disabled people. Comparing with other biometric features, EEG has several advantages as follows: (a) it is confidential (as it corresponds to a mental task), (b) it is very difficult to mimic (as similar mental tasks are person dependent), (c) it is almost impossible to steal (as the brain activity is sensitive to the stress and the mood of the person, an aggressor cannot force the person to reproduce his/her mental pass-phrase). [3] In this patent we are building a concept of brain print and assuming that EEG signal alone is able to create a unique pattern for each subject. In other words we are not going to combine any other human feature with EEG signal to identify people. We are considering working on large number of peoples with two public databases and using simple feature extraction and simple classification methods to provide strong evidence that our novel algorithm with EEG signal processing can provide unique patterns to identify people with other human features. [4] With a data set of four subjects and 255 EEG trials (subjects were at first with eyes closed) Poulos et al. adopted two classification algorithms and obtained the accuracies of around 80% and 95% respectively. Paranjape et al. analysed a data set of 40 subjects and 349 EEG trials (subjects were resting with eyes open and closed) and got a classification accuracy of about 80%. Palaniappan and Mandic carried out a personal identification experiment with 102 subjects based on visual evoked potentials and the accuracies were around 95-98%. Marcel and Mill'an got a highest accuracy rate for personal verification of 93.4%. [5] The above early work has played an important role in studying the feasibility of EEG signals for usage in biometrics. However, when learning a classifier, they all adopted only one kind of brain activity. [6] In order to make human identification system more effective and efficient, we particularly focus on the simplest algorithm for decreasing the calculations and shorten latency. [7] In this work we are trying to test the EEG uniqueness over a large number of subjects, and also try to use simpler method for feature extraction to make EEG identification more applicable. [8] So in this work we shall: (1). Use EEG identification method on a large number of subjects to emphasize EEG uniqueness among peoples. This will enhance the opportunity to use EEG identification on large scale, or even to use it as a universal human identity. (2). Use relatively low complexity and low computation cost methods in pre-processing and feature extraction, to enhance considering EEG as an online solution for human identification. In this work we tackled the above concerns by using large public database that contains EEG data for (i) 45 people and (ii) 122 people. Also all the processing are only considering rms spatial pattern only to create feature vector which is used for the first time in EEG.

2 [9] There are many debates about EEG bandwidth and it is noted that significant signals are distributed within lower than 100 Hz, for example Howard et al., where they suggest upper limit to gamma in EEG bandwidth to 60Hz. A typical set of EEG signal during a few seconds for an adult brain activity are as shown in Figure 1. Therefore, in the pre-processing step, all the EEG signals were filtered to get frequencies between 0 and 60 Hz. All frequency components above 60 Hz were disregarded. There shows an example of the effect of the filtration on of the EEG signals. For the extraction of the human feature, the whole processing will only take EEG low pass signals and mapping them into the rms value for each special position or each electrode and the rms value represents active potential of the signal where power of the signal p(x) is directly proportional to the rms value. [10] We have designed the EEG sample from each electrode is divided into one second time period length signals including 256 values, and the rms values for all the 256 values are calculated with equation (1) then sending them to feature vector. For this case we obtained feature vectors of length 64 rms values that taken from the related electrodes as shown by Figure 1. [11] For the extraction of the human feature, the whole processing will only take EEG low pass signals and mapping them into the rms value for each special position or each electrode. Each rms value, denoted as x, can be obtained from the well know definition of equation (1): n and the rms value represents active potential of the signal where power of the signal p(x) is directly proportional to the rms value, as shown below: p(x) oc rms 2 (x) (2) [12] A neural network (NN) classifier is designed to classify the obtained data. The NN classifier is feed forward error back propagation network. Training starts from a random weight set. The NN is designed with 64 nodes in the input layer, which is the same number of electrodes. The number of outputs depends on the number of subjects which is 45 for the first experiment and 122 for the second experiment. The network has 45 neurons hidden layer in the first experiment, and 70 neurons hidden layer in the second experiment. In the second experiment which was operated on 122 subjects. We used the MATLAB built in nntraintool tool to run the tests. The rms feature vector input was pre-processed by this tool by normalizing the data between [1, -1]. [13] As mentioned in above that the dataset was taken from the public data repository for machine learning. This dataset was collected through a study was performed at the Neurodynamics Laboratory of the State University of the New York Health Centre at Brooklyn. This study EEG correlates of genetic predisposition to alcoholism. The dataset contains multiple measurements from 64 electrodes placed on subject's scalps which were sampled at 256 Hz (3.9 ms epoch) for 1 second. [14] There were two groups of subjects: alcoholic and control. Each subject was exposed to either a single stimulus (S1) or to two stimuli (S1 and S2) which were pictures of objects chosen from the 1980 Snodgrass and Vanderwart picture set. When two stimuli were shown, they were presented in either a matched condition where S1 was identical to S2 or in a non-matched condition where S1 differed from 3 S2. There were 122 subjects and each subject completed 120 trials where different stimuli were shown. Zhang et al. (1995) describes in detail the data collection process. [15] The original data contains 77 alcoholic subjects and 45 control subjects. In the first experiment we consider half the samples available for all 45 control subjects. [16] The samples were selected randomly. The input layer size is 64 inputs which is the number of rms value for each electrode. The NN back propagation with one hidden layer with a number of neurons equal to the number of outputs (45), and the output layer which represent the number of subjects (45 control peoples). The NN engine by default normalizes the data between 1 and -1 for the input and output. The training stopped when the classifier reached below the minimum gradient which is set to 10-6. [17] Obviously, the results were so promising, and the classifier was able to identify all the 45 peoples correctly, with a mean square error value of 1.98842 x10- 7 . The similar design was used to the second dataset. The target is trying to check if this algorithm has generalization for dealing with other EEG signal. The second database is about 122 people in comparison the first database the size is almost three times as the previous one, which is obvious a good challenge to the designed algorithm. Also this will verify if the rms spatial pattern can be considered as a brain signature or brain print. In the second experiment the input size remain the same which 64 rms inputs for the EEG electrodes. The hidden layer size was increased to be 70 neurons arbitrarily. And the output size is 122 which is the number of peoples. As in the first experiment the NN engine by default normalizes the data between 1 and -1, and the continuous tan sigmoid activation function was used. [18] Although we consider bigger number of peoples, the results was also promising. The classifier was able to identify 113 peoples correctly out of 122, with a mean square error value of 0.00271. The other nine subjects were clarified the case that the classifier was not able to identify: four of nine were highly confused with other subject in the sample, and five were not identified totally. Figure 2 shows the mean square error and Figure 3 shows the gradient during the training. [19] To enhance the efficiency of the classifier in the second experiment, we add a weighted connection between the input layer and the output layer. The efficiency increases after this enhancement, and the classifier was able to identify 116 peoples correctly out of 122, in other words 95.1% successful rate. [20] The mean square error value was 0.00186. The other six subject that the classifier were not able to identify, four of them were highly confused with other subject in the sample, and these four are different than the four in the first part of this experiment. The other two were not identified totally. This last experiment shows that by enhancing the classifier the result might enhance and a better classification rate might be achieved through using the rms spatial pattern as a feature vector. Figure 4 shows the mean square error and Figure 5 shows the gradient during the training of this experiment. [21] In this paper we have been focusing on one of non-invasive brain computer interface (BCI) signal, a typical variety of brain signals, electroencephalography (EEG) as input to analysis its characteristics. Those characteristics are used to identify the people as other biometrics to recognize and distinguish people based on their physical or behavioral features. As using EEG signals to identify people has some advantages such as it is confidential, it is very difficult to mimic and it is almost impossible to steal, etc. EEG signal processing has drawn great attentions as this paper does. A novel algorithm is 4 presented in this paper. Our designed classifier is able to identify all the 45 people correctly with a mean square error of 2.03 34E 10-7 for the first public open database and the same algorithm applying to the 2nd database with 116 out of 122 people can be fully identified (successful rate of 95.1%) with a mean square error value of 0.00186. [22] We deeply believe that a low complexity, high resolution, effective and efficient is very attractive for the real life applications become true in the foreseeable future.

Claims (6)

1. A method is using electroencephalogram (EEG) signal to identify human through the root mean square (rms) values of obtained signals.

2. A method according to claim 1, the EEG bandwidth is focus on the spectrum between 8 to 57 Hz signals contributes to identify human.

3. A method according to claim 1, the network has 45 neurons hidden layer in the first experiment, and 70 neurons hidden layer in the second experiment. In the second experiment which was operated on 122 subjects. We used the method that MATLAB built in nntraintool tool to run the tests with proposed rms values.

4. A method according to claims 1 and 3, the rms feature vector input was pre-processed by this tool by normalizing the data between [1, -1], which has been used for the network identification.

5. A method according to claim 4, the dataset contains multiple measurements from 64 electrodes placed on subject's scalps which were sampled at 256 Hz (3.9 ms epoch) for 1 second.

6. A method of recoding in the second experiment, the input size remains the same which 64 rms inputs for the EEG electrodes. The hidden layer size was increased to be 70 neurons arbitrarily.