ECG Identity Authentication in Open-set with Multi-model Pretraining and Self-constraint Center & Irrelevant Sample Repulsion Learning

TL;DR

This paper tackles open-set ECG identity authentication by introducing a two-stage framework: multimodal pretraining that aligns ECG signals with fiducial-feature text reports to enrich representations, and a fine-tuning regime with Self-constraint Center Learning and Irrelevant Sample Repulsion Learning to enforce compact, discriminative feature boundaries. The approach leverages class centers, dynamic prototypes, and reciprocal points to tighten intra-class clustering while pushing open-set distributions away from the closed-set space, demonstrated through substantial improvements in ACC and OSCR and a low FAR across multiple open-set scenarios. Key results show ACC reaching 99.83% on closed-set data and FAR as low as 5.39% with OSCR consistently above 95% as open-set proportion varies, outperforming several strong baselines on ECGID, MIT-BIH, and Autonomic datasets. The work provides a practical pathway to robust ECG-based authentication in open-world settings, though it acknowledges limitations under very large open-set regimes and suggests future work to further reduce false acceptances.

Abstract

Electrocardiogram (ECG) signal exhibits inherent uniqueness, making it a promising biometric modality for identity authentication. As a result, ECG authentication has gained increasing attention in recent years. However, most existing methods focus primarily on improving authentication accuracy within closed-set settings, with limited research addressing the challenges posed by open-set scenarios. In real-world applications, identity authentication systems often encounter a substantial amount of unseen data, leading to potential security vulnerabilities and performance degradation. To address this issue, we propose a robust ECG identity authentication system that maintains high performance even in open-set settings. Firstly, we employ a multi-modal pretraining framework, where ECG signals are paired with textual reports derived from their corresponding fiducial features to enhance the representational capacity of the signal encoder. During fine-tuning, we introduce Self-constraint Center Learning and Irrelevant Sample Repulsion Learning to constrain the feature distribution, ensuring that the encoded representations exhibit clear decision boundaries for classification. Our method achieves 99.83% authentication accuracy and maintains a False Accept Rate as low as 5.39% in the presence of open-set samples. Furthermore, across various open-set ratios, our method demonstrates exceptional stability, maintaining an Open-set Classification Rate above 95%.

Figures (4)

• Figure 1: The proposed method is outlined in the workflow diagram, which consists of two main components: multi-modal pretraining and fine-tuning on identity authentication task. Within component B, we introduce Self-constraint Center Learning and Irrelevant Sample Repulsion Learning.
• Figure 2: The experimental results comparing various baseline methods are presented using ACC, OSCR, FAR, TNR, and AUC.
• Figure 3: Line chart illustrating the variations in OSCR, FAR, and TNR as the ratio of open-set data to close-set data changes.
• Figure 4: T-SNE visualizations of sample feature distributions under different ablation settings. Notably, (b) and (f) represent the results obtained using only the irrelevant sample repulsion learning module (only with B.1). (c) and (g) represent the results obtained using only the self-constraint center learning module (only with B.2). (d) and (h) represent the results obtained using only the dynamic prototype learning learning module (only with B.3).