2011 International Conference on Image Information Processing (ICIIP 2011) Proceedings of the 2011 International Conference on Image Information Processing (ICIIP 2011) 978-1-61284-861-7/11/$26.00 ©2011 IEEE ECG based Biometric Authentication-a Novel Data Modelling Approach Saurabh Pal Dept. of Applied Elec. and Instrumentation Engineering Heritage Institute of Technology Kolkata, India Madhuchhanda Mitra Dept. of Applied Physics University of Calcutta Kolkata, India Abstract— Use of ECG for biometric authentication is a relatively new approach for in security and restricted access methodologies. Here an ECG based technique for biometric analysis is proposed for a group of 20 persons. The method is based on first accurate extraction of characteristic features from each ECG and then design of a classifier for authentication. The feature data set is dimensionally reduced before classification. Classification is based on a simple comparison of the signature matrix elements for stored and testing data. In this work it is shown that the classification accuracy greatly improves if the reduced data set is modelled by a quadratic polynomial based on least square. Keywords- biometric authentication, ecg, pca, quadratic polynomial. I. INTRODUCTION Biometric authentication is a well adopted technique for security purposes. Most common biometric parameters include finger print, iris pattern, face and voice recognition etc as anatomical and physiological feature or signature, key stroke dynamics etc are the behavioural features. However, these biometrics modalities either cannot provide reliable performance in terms of recognition accuracy (e.g., gait, keystroke) or they are prone to falsification externally. For an example, fingerprint or face can be modified by plastic surgery, voice can be modified by DSP based techniques or iris features can be changed by contact lenses with some other features printed on it. In this regard, ECG has a good potential which can be used alone as a biometric parameter or in combination with some other parameter for greater accuracy. It has the advantage that the source of ECG signal is heart which is not easily accessible externally neither it is easy to modify the signal. Moreover, ECG signal can give the liveliness proof which is not possible to provide by the other biometric parameters. Recently some work on the applicability of ECG in biometry is reported by different researchers. Biel et al. [1] have conducted some experiments on ECG based biometry for a group of 20 subjects. Twelve features have been selected from each record for identification of a person in a predefined group. Shen et al. [2] have investigated the feasibility of ECG as a new biometric for identity verification. The experiment has been conducted on 20 individuals on seven features, extracted mainly from QRS complex. T.W.Shen has also shown [3] that it is possible to identify people with a one-lead ECG signal on a small group. There are two methods has been investigated for population (<30). Wang et. al. [4] has proposed methods using fiducial points and without fiducial points by Autocorrelation/Discrete Cosine Transform (AC/DCT) technique for two groups with 13 subjects in each. Classification is done based on linear discrimination analysis and neural network based technique. In AC/DCT method similarity between the subjects is measured based on normalized Euclidian distance and a nearest neighbour is used as the classifier. Chan et. al. [5] use a wavelet distance measurement technique for classification of 50 subjects with accuracy 89%. Some of the previous studies are based on ECG template matching which does not require the detection of ECG characteristic points. This technique has the disadvantage that the test fails if two persons have similar ECG. The method s involving fiducial point detection suffers problem of handling a large number of dataset. Thus it requires database reduction keeping the discriminating features intact. In this work the data is dimensionally reduced by Principal Component Analysis and the reduced feature set is fitted into a quadratic polynomial model for better classification as then the higher order elements of the feature matrix shows more difference within the data set. Finally the decision of acceptance or rejection is made by comparison. The block diagram of the procedure is shown in figure 1. Figure 1. Bolck diagram of the entire procedure II. MATERIALS AND METHODS A. Feature Extraction In this work wavelet transform based feature extraction is performed as described in [6].Wavelet transform is basically a convolution operation between the mother wavelet and the test