Adaptive Convolutional Neural Network for Robust ECG-based Biometric Authentication using Variable-length Signals

Abstract

Electrocardiogram (ECG) signals have emerged as a reliable biometric modality due to their intrinsic physiological origin, inherent liveness detection capability, and resistance to forgery attacks. However, conventional ECG-based authentication systems typically rely on fixed-length heartbeat segmentation, which limits their robustness under natural variations in cardiac cycle duration and reduces authentication reliability. To address this limitation, this paper proposes an improved convolutional neural network (CNN) architecture designed to perform biometric authentication directly from variable-length ECG signals without enforcing strict input normalization. The proposed model enables automatic extraction of discriminative identity-specific features while preserving the natural temporal variability of ECG waveforms, thereby enhancing robustness and generalization across subjects. The proposed framework was evaluated using the publicly available ECG-ID dataset, which contains ECG recordings collected under realistic conditions from multiple individuals. Experimental results demonstrate that the proposed variable-input CNN achieves an authentication accuracy of 97.6%, outperforming conventional fixed-input CNN approaches in terms of stability and robustness. The novelty of this work lies in the development of a flexible deep learning-based authentication framework capable of handling variable-length ECG inputs, improving real-world applicability and reliability of physiological biometric systems for secure authentication applications