심전도를 이용한 LSTM과 2D-CNN의 앙상블기반 개인식별

Abstract

Recently, with the rapid development of artificial intelligence, technology has been applied to various fields to improve the quality of life and work, not only in human daily life, but also in medical care, security, and finance. However, there are cases where humans are threatened with security problems by exploiting this, and technologies that can authenticate only various individuals are expanding in order to defend against this. Previously, personal information such as passwords and OTP was protected, but there is a problem of loss or theft, so personal biometric information can be used to supplement important personal information. Biometrics is a physical and behavioral feature, and personal authentication technology using biometric information requires characteristics of universality, uniqueness, permanence, collectability, accuracy, acceptability, and circumvention.

In this paper, an individual was identified by using an electrocardiogram(ECG) signal with the characteristics of excellent identification. A low-pass filter, a high-pass filter, and an average shift filter were applied to the signal as a preprocessing process, and the baseline was adjusted to zero. Also, the signal was divided at regular intervals based on the R-peak. The performance of the Long Short-Term Memory(LSTM) neural network consisting of one LSTM layer and two LSTM layers was compared and analyzed. In addition, the ECG signal was converted into the time-frequency domain for signal analysis, and expressed as two-dimensional images of Short-Time Fourier Transform(STFT), Scalogram, Fourier Synchrosqueezed Transform(FSST), and Wavelet Synchrosqueezed Transform(WSST). Deep models od CNN were used by GoogleNet, VGG-19, and ResNet-101. In this paper, the used database consists of the Chosun University(CU)-ECG database and the Physikalisch-Technische Bundesanstalt(PTB) database. Finally, the experimental results showed high performance from a minimum of 1.06% to a maximum of 3.75% and from a minimum of 0.8% to a maximum of 3.38% for CU-ECG databases and PTB database, respectively.