Out-of-Distribution Representation and Graph Neural Network Fusion Learning for ECG Biometrics

The electrocardiogram (ECG) signal, a promising trait in biometrics, has been extensively studied. While the deep learning-based model has demonstrated strong performance for ECG biometrics, several challenges remain, including efficient extraction of 1D signals’ topological properties and efficient use of signal distribution information at different times. Another challenge is adapting to the critical role of specific deep neural networks. To address these issues, this study proposes an out-of-distribution representation and graph neural network fusion learning (ORGNNFL) method for ECG biometrics. The ORGNNFL is mainly composed of a two-branch deep learning model for ECG biometrics, capable of learning to discriminate features from latent distributions and extracting topological features of 1D ECG signals. The multi-feature attention module is also proposed to adaptively capture valuable information from two-branch deep learning data for ECG biometrics. Experiments conducted on the four databases demonstrate that our method outperforms the state-of-the-art techniques, paving the way for a new era in ECG biometrics. Code is available at https://github.com/matianbang/ORGNNFL.

I. Introduction

B iometric recognition is the widespread field that models people’s recognition using their physiological or behavioral traits, such as face [1], fingerprint [2], iris [3], voiceprint [4], gait [5], and handwriting [6]. In the last decade, the individual-unique information embedded in bioelectric signals has been gradually mined, and ECG biosignals have been investigated as potential biometric traits [7]. In contrast to external biometric traits, ECG signals have several distinctive characteristics [8]: 1) Liveness detection. ECG signals are recorded by sensors attached to the body, and they can only be captured from a living person. 2) High privacy. ECG signals are difficult to fake at the acquisition stage. 3) Hybrid information. ECG signals contain rich information that can be applied in diverse areas including identity recognition, heart disease monitoring, and emotion analysis. 4) Convenient storage. ECG signals are one-dimensional data that are easy to store and process.