A Multi-Scale CNN–BiLSTM Framework for Robust ECG-Based User Authentication

Abstract

Reliable biometric authentication remains a critical challenge for modern security systems, particularly in applications requiring continuous verification and strong resistance to spoofing attacks. Among physiological biometrics, the electrocardiogram (ECG) offers inherent liveness information, subject-specific morphological patterns, and robustness against external forgery. This paper presents a novel deep learning architecture for ECG-based user authentication that jointly models spatial morphology and temporal dynamics through an integrated multi-scale convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) framework. The proposed model uses parallel convolutional branches with different kernel sizes to simultaneously capture discriminative characteristics of the P-wave, QRS complex, and T-wave at multiple temporal scales. The extracted multi-scale features are fused and subsequently processed by a BiLSTM layer to exploit both forward and backward temporal dependencies across heartbeat sequences. Extensive experiments are conducted on the publicly available PTB-XL dataset using a subject-independent evaluation protocol. The proposed approach achieves an authentication accuracy of 99.2% and an equal error rate of 0.8%, outperforming conventional CNN, LSTM, and unidirectional CNN–LSTM baselines across all evaluation metrics. The findings indicate a noteworthy aspect of integrating multi-scale morphological research with bidirectional temporal modeling. This combination significantly enhances ECG-based biometric authentication. Enhances its robustness, hence reducing the likelihood of failure in atypical conditions. Moreover, reliability increases, which is essential for security-related matters such as this