ECG Biometric Authentication Using Self-Supervised Learning for IoT Edge Sensors.

Area of Science:

• Computer Science
• Biomedical Engineering
• Cybersecurity

Background:

• Wearable Internet of Things (IoT) devices enable continuous physiological data collection for health monitoring.
• Physiological signals, such as electrocardiograms (ECG), offer potential for passive, continuous user authentication in security applications.
• Existing authentication methods often lack convenience and continuous security.

Purpose of the Study:

• To investigate an ECG-based biometric user authentication system leveraging Convolutional Neural Networks (CNN) and self-supervised contrastive learning.
• To develop a robust system for continuous authentication in wearable IoT devices.
• To optimize the model for deployment on resource-constrained embedded devices.

Main Methods:

• Utilized a CNN architecture combined with self-supervised contrastive learning to extract distinguishing ECG features.
• Trained and evaluated the model on the PTB ECG database (290 subjects).
• Assessed model generalizability on the MIT-BIH Arrhythmia and ECG-ID databases.
• Applied model optimization techniques including quantization and pruning for embedded deployment.

Main Results:

• Achieved 99.15% authentication accuracy on the PTB ECG database.
• Demonstrated high generalizability with over 98.5% accuracy on unseen datasets (MIT-BIH Arrhythmia, ECG-ID) without retraining.
• Repeating authentication thrice increased accuracy to nearly 100% on PTBDB and ECGIDDB.
• Optimized model retained 98.67% accuracy on PTBDB while reducing CPU cycles by 62.6%.

Conclusions:

• ECG-based biometrics with CNN and contrastive learning provide a highly accurate and generalizable solution for continuous user authentication.
• The proposed system is suitable for real-world deployment on wearable IoT devices, even after optimization for embedded systems.
• Model optimization techniques effectively balance accuracy and computational efficiency for IoT edge sensors.