HMM BASED BIOMETRIC SYSTEM USING CARDIAC SIGNALS

Abstract

A pattern recognition system which is able to recognize a user is essentially referred to as a biometric system. In this paper, two types of biometric signals were used to build the proposed multimodal biometric system; the Electrocardiogram (ECG) and Heart Sound (HS). The ECG and HS data are not commonly used as biometric due to the signal characteristic complexity which make it very hard to duplicate and more immune to spoof attacks. This work was conducted for Client Identification (CID) with fixed 20 clients, the data were sampled at 44 kHz for the two biometric signal. An adaptive windowed approach of Mel Frequency Cepstral Coefficients (MFCC) was used to extract the features. The extracted features then partitioned into train and test sets, the train set fed to Hidden Markov Model (HMM) to create the independent-client trained model. The purposed biometrics system is based on the performance of two folds of training sets, 30% and 70%. Complexity of states and Gaussians also plays a role on the performance. The best performance for CID with 44 kHz, evaluated with 20 clients is based on HS which provide an accuracy of 93.04% with training data of 70%. The worst performance goes to 87.89% for ECG at 30%.