Off-Person ECG Biometrics Using Spatial Representations and Convolutional Neural Networks

Abstract

Targeting off-person ECG-based biometrics, we report a comparative analysis of identification accuracy and verification Equal Error Rate (EER) performances of four distinct types of spatial representations of ECG signals applied as inputs to Convolutional Neural Networks. The actual algorithms used to transform the original time series into 2D/3D images are based on a modified version of the Continuous Wavelet Transform (the S-Transform), the Gramian Angular Field, the recurrence plot, and state-space representations. Extensive experiments have been conducted using UofT and CYBHI datasets including recordings acquired on fingers and hand palm under various activity scenarios. The wavelet-based approach yielded best results, while all analyzed solutions compare favorably with previously reported performances.

Keywords

• Biometrics
• off-person
• ECG
• spatial representation
• deep learning