Biometric Authentication Using the Unique Characteristics of the ECG Signal

Abstract

ECG is a biological signal specific for each person that is hard to create artificially. Therefore, its usage in biometry is highly investigated. It may be assumed that in the future, ECG for biometric purposes will be measured by wearable devices. Therefore, the quality of the acquired data will be worse compared to ambulatory ECG. In this study, we proposed and tested three different ECG-based authentication methods on data measured by Maxim Integrated wristband. Specifically, 29 participants were involved. The first method extracted 22 time-domain features – intervals and amplitudes from each heartbeat and Hjorth descriptors of an average heartbeat. The second method used 320 features extracted from the wavelet domain. For both methods a random forest was used as a classifier. The deep learning method was selected as the third method. Specifically, the 1D convolutional neural network with embedded feed-forward neural network was used to classify the raw signal of every heartbeat. The first method reached an average false acceptance rate (FAR) 7.11% and false rejection rate (FRR) 6.49%. The second method reached FAR 6.96% and FRR 21.61%. The third method reached FAR 0.57% and FRR 0.00%.ECG is a biological signal specific for each person that is hard to create artificially. Therefore, its usage in biometry is highly investigated. It may be assumed that in the future, ECG for biometric purposes will be measured by wearable devices. Therefore, the quality of the acquired data will be worse compared to ambulatory ECG. In this study, we proposed and tested three different ECG-based authentication methods on data measured by Maxim Integrated wristband. Specifically, 29 participants were involved. The first method extracted 22 time-domain features – intervals and amplitudes from each heartbeat and Hjorth descriptors of an average heartbeat. The second method used 320 features extracted from the wavelet domain. For both methods a random forest was used as a classifier. The deep learning method was selected as the third method. Specifically, the 1D convolutional neural network with embedded feed-forward neural network was used to classify the raw signal of every heartbeat. The first method reached an average false acceptance rate (FAR) 7.11% and false rejection rate (FRR) 6.49%. The second method reached FAR 6.96% and FRR 21.61%. The third method reached FAR 0.57% and FRR 0.00%.