Abstract
To enhance the security level of digital information, the biometric authentication method based on Electrocardiographic (ECG) is gaining increasing attention in a wide range of applications. Compared with other biometric features, e.g., fingerprint and face, the ECG signals have several advantages, such as higher security, simpler acquisition, liveness detection, and health information. Therefore, various methods for ECG-based authentication have been proposed. However, the generalization ability of these methods is limited because the feature extraction for the ECG signals in conventional methods is data dependent. To improve the generalization ability and achieve more stable results on different datasets, a parallel multi-scale one-dimensional residual network is proposed in this paper. This network utilizes three convolutional kernels with different kernel sizes, achieving better classification accuracy than the conventional schemes. Moreover, two loss functions named center loss and margin loss are used during the training of the network. Compared with the conventional softmax loss, these two loss functions can further improve the generalization ability of the extracted embedding features. Furthermore, we evaluate the effectiveness of our proposed method thoroughly on the ECG-ID database, the PTB Diagnostic ECG database, and the MIT-BIH Arrhythmia database, achieving 2.00%, 0.59%, and 4.74% of equal error rate (EER), respectively. Compared with other works, our proposed method improves 1.61% and 4.89% classification accuracy on the ECG-ID database and the MIT-BIH Arrhythmia database, respectively.
Introduction
While fingerprints and facial recognition are standard, they are "external" biometrics that can be spoofed or captured without consent. Electrocardiogram (ECG) signals are "internal," providing inherent liveness detection—you cannot authenticate without a pulse—and additional health data.
However, ECG-based authentication faces a major hurdle: generalization. Most existing models perform well on the specific data they were trained on but struggle when faced with different recording environments or "unseen" individuals. Traditional "fiducial" methods, which rely on measuring specific peaks like the P-Q-R-S-T waves, are also notoriously sensitive to noise and require manual, error-prone feature selection. This paper presents a robust approach to biometric authentication using a multi-scale 1D Residual Network combined with a sophisticated triple-loss training strategy to overcome these limitations.
Proposed Methodology
The proposed system follows a structured pipeline designed to transform raw electrical heart signals into highly discriminative identity "embeddings."
1. Preprocessing
Raw ECG signals are often contaminated by noise, such as powerline interference and muscle tremors. The preprocessing stage ensures the data is clean and standardized for the neural network.
As shown above, the signals are resampled to 500 Hz and filtered. The system then detects R-peaks and extracts 300-point segments around each heartbeat to create consistent inputs for the model.
2. Parallel Multi-Scale 1D ResNet
Instead of a single chain of layers, the model uses three parallel branches with different convolutional kernel sizes ($1 \times 3$, $1 \times 5$, and $1 \times 7$). This allows the network to simultaneously capture fine-grained details (like the sharp R-spike) and broader morphological patterns (like the T-wave) of the heartbeat. Residual "skip" connections are integrated to prevent vanishing gradients and improve training efficiency.
3. The Triple Loss Strategy
To make the resulting feature vectors (embeddings) more robust, the authors utilized three distinct loss functions during training:
- Softmax Loss: Ensures basic class separability.
- Center Loss: Pulls embeddings of the same person toward a central point, increasing "intra-class compactness."
- Margin Loss: Enforces a minimum distance between different people's embeddings, increasing "inter-class dispersion."
Experimental Results
The model was pre-trained on the massive PhysioNet/Cinc Challenge 2017 dataset and then evaluated on three distinct benchmarks to test its generalization capabilities:
- ECG-ID Database: 90 healthy subjects; achieved 2.00% EER and 98.24% accuracy.
- PTB Diagnostic ECG Database: 52 healthy subjects; achieved 0.59% EER and 100% accuracy.
- MIT-BIH Arrhythmia Database: 47 subjects with heart disease; achieved 4.74% EER.
The performance on the MIT-BIH database is particularly noteworthy, as it demonstrates the model's ability to maintain high accuracy even when heartbeats are irregular due to underlying pathology.
Conclusion
The proposed parallel multi-scale 1D ResNet significantly improves the generalization ability of ECG-based authentication. By shaping the feature space with Center and Margin losses, the system creates robust embeddings that outperform conventional methods. Future work may involve optimizing these networks for low-power hardware accelerators suitable for wearable devices and exploring the impact of physiological stressors like exercise on authentication performance.
Summary updated