Deep Neural Networks for Biometric Identification Based on Non-Intrusive ECG Acquisitions

Abstract

The electrocardiogram has gained traction as a biometric trait due to its outstanding combination of universality, permanence, and measurability, with a hidden nature that makes it harder to steal or counterfeit. The state-of-the-art mostly consists of pipeline algorithms, composed of separate stages of denoising, segmentation, feature extraction, and decision. However, Convolutional Neural Networks (CNNs), possess the tools to integrate all phases of processing, from acquisition to decision, in a single model. This integration replaces separate, step-by-step tuning with an holistic optimisation process, synergically adapting the model to attain the best performance possible. In this chapter, we introduce and explore the capabilities of convolutional neural networks for biometric identification using non-intrusive ECG signal acquisitions, and propose a CNN architecture for the complete integration of traditional pipeline stages in a single accurate and robust model. The method was evaluated on the highly complete and challenging UofTDB collection, and has shown promising results on identification tasks when compared with recent and successful state-of-the-art methods.