Comparative Evaluation of Nonfiducial Techniques Toward User Authentication by Noncontact Electromagnetic Sensor on Chest Wall

We evaluate user authentication using chest wall displacement as time-series signals recorded by a wearable near-field electromagnetic sensor (NFS) and nonfiducial signal processing techniques. Although the NFS can record the change in the input impedance of an antenna as a time-series NFS signal based on the near-field interaction between the chest wall and the antenna with minimal restrictions, definitive fiducial features and their extraction for the signal waveforms have not been established. Therefore, we recorded NFS signals from 22 participants using an implemented NFS module and compared traditional and deep learning-based models containing nonfiducial feature extraction with two segmentation methods applied to the signals to achieve optimal performance, considering the comparison and combinations of nonfiducial and fiducial features from NFS I/Q signals. As a result, a deep learning-based model combining convolutional layers and long short-term memory with nonfiducial fixed length-based segmentation utilizing both I and Q signals achieved an equal error rate of 0.033 $\pm$ 0.030.