LIGHTWEIGHT SELF-SUPERVISED ECG AUTHENTICATION FOR RESOURCE-CONSTRAINED IOT EDGE SENSORS

Azab Mohamed AbdallaElsayed, Sila Anastasiia Stanislavovna, Korzhuk Viktoriia Mikhailovna

This paper proposes a lightweight ECG biometric authentication framework for wearable IoT devices using CNN-based feature extraction and self-supervised contrastive learning. The approach enables effective training using unlabeled ECG signals and improves cross-dataset generalization. The model achieved 99.15% accuracy on the PTB dataset and maintained over 98.5% accuracy on MIT-BIH and ECG-ID datasets without retraining. Model optimization through pruning and quantization reduced computational requirements with minimal performance loss, achieving 98.67% accuracy. The findings demonstrate the feasibility of deploying robust ECG authentication on resource-limited IoT edge platforms.

Keywords:Electrocardiogram Authentication, Contrastive Learning, IoT Devices.