TLDR: Researchers propose a lightweight deep learning model, MobileNetV1+GRU, for secure ECG-based biometric authentication on wearable devices. The model achieves high accuracy (91.74%–99.34%) across four diverse datasets by employing custom preprocessing, including 20dB Gaussian noise injection and R-peak segmentation. It demonstrates promising robustness and generalization, with evaluations covering federated learning for privacy and adversarial attacks to assess vulnerabilities, outperforming existing methods.
Electrocardiogram (ECG) signals, unique to each individual, are emerging as a highly secure method for biometric authentication, especially for wearable devices. Unlike traditional biometrics such as fingerprints or facial recognition, ECG patterns are difficult to spoof or alter, offering a robust solution for personal identification. However, deploying ECG-based authentication on wearable technology presents several challenges, including real-time processing demands, privacy concerns, and vulnerability to sophisticated attacks.
A recent research paper introduces a novel, lightweight deep learning model that aims to overcome these hurdles: the MobileNetV1+GRU architecture. This innovative model is specifically designed for ECG-based authentication, making it suitable for resource-constrained wearable devices. The researchers focused on simulating real-world conditions by injecting 20dB Gaussian noise into ECG signals and implementing custom preprocessing techniques to enhance the model’s robustness.
The MobileNetV1+GRU Model: A Hybrid Approach
The proposed system combines MobileNetV1, a convolutional neural network known for its efficiency in spatial feature extraction, with a Gated Recurrent Unit (GRU), which excels at processing sequential data. This hybrid approach allows the model to effectively analyze the complex spatial and temporal patterns inherent in ECG waveforms. The preprocessing steps are crucial: ECG signals are bandpass filtered, injected with noise, and normalized. A key innovation is the precise R-peak detection and segmentation, where heartbeats are isolated and centered around the R-peak, creating standardized segments for analysis. These segmented beats are then transformed into 224×224 CWT–Morlet scalograms, which are visual representations of the ECG signal’s frequency content over time, providing rich features for the MobileNetV1 component.
Rigorous Evaluation Across Diverse Datasets
To ensure broad applicability and generalization, the model was extensively evaluated using four distinct ECG datasets: ECG-ID, MIT-BIH, CYBHi, and PTB. These datasets represent a variety of subjects and acquisition conditions. The model demonstrated impressive performance, achieving accuracies of 99.34% on ECG-ID, 99.31% on MIT-BIH, 91.74% on CYBHi, and 98.49% on PTB. Other key metrics like F1-scores, precision, recall, equal error rates (EER), and ROC-AUC values also indicated high effectiveness, with ROC-AUC values consistently near 0.9999 for most datasets.
Addressing Security and Privacy: Federated Learning and Adversarial Attacks
The paper also delves into critical aspects of security and privacy. It explores federated learning (FedAvg), a distributed machine learning approach that allows models to be trained on decentralized data without sharing raw information, thereby preserving user privacy. The evaluation showed that federated learning effectively aggregates client models, offering a promising path for privacy-preserving biometric systems. Furthermore, the model’s resilience against adversarial attacks was tested using the Fast Gradient Sign Method (FGSM). While accuracy dropped significantly under strong attacks (e.g., from 96.82% to as low as 0.80% for ECG-ID under a high epsilon attack), the study highlights the importance of dataset quality and training strategies in building robust models, especially noting that models trained on ECG-ID, PTB, and MIT-BIH were more robust than those trained on CYBHi, which was collected under less controlled conditions.
Also Read:
- ECG Data Alone Can Accurately Identify Human Activities, New Study Shows
- SelfMIS: Enhancing Heart Attack Detection from Single-Lead ECGs Through Latent Space Alignment
Interpretability and Future Directions
The research also incorporated Explainable AI (XAI) techniques, such as EER curves and occlusion-based Integrated Gradients (IG) highlighting, to provide insights into the model’s decision-making process. This interpretability is vital for trust and understanding in biometric systems. The proposed MobileNetV1+GRU model not only outperformed existing methods in ECG biometric authentication but also demonstrated its potential for broader applications like cardiac monitoring and heart condition detection. Future work aims to explore transformer-based models, multi-factor authentication, and advanced preprocessing techniques to further enhance spoofing resistance and robustness in diverse physiological conditions. For more in-depth information, you can refer to the full research paper available here.
