ECG Signal-Based User Identification with Deep Learning: Leveraging Morphological Characteristics and Domain Adaptation

Traditional biometric systems are increasingly vulnerable to spoofing attacks due to their lack of liveness detection. Electrocardiogram (ECG) signals offer a viable biometric trait that combines inherent liveness and robust spoof resistance, making it a promising alternative for user identification. However, physiological variations of ECG signals pose significant pattern recognition challenges. Current research has demonstrated high accuracy in analyzing subjects in resting states, but performance significantly decreases when evaluating subjects in non-resting states as we demonstrate with a dataset that contains post-exercise data.
We examine a fundamental scenario involving multi-session training data in resting states, which results in high baseline accuracy for rest state evaluation data. Adapting existing methods, typically involving training data from a single session, to our multi-session scenario, led to a notable decline in performance. While post-exercise recognition improved, the error rate for rest state evaluation doubled. In order to overcome this decline in performance, we propose a tailored deep convolutional neural network (CNN) architecture designed to leverage known morphological characteristics of ECG signals, coupled with personalized augmentation and domain adaptation of the classifier.
The proposed personalized augmentation technique modifies the ST segment within constrained ranges predicted individually based on resting data, leading to enhanced augmentation efficacy. This approach deviates from previous studies that assume a uniform pattern of ST segment change for all subjects.
Domain adaptation is used to address the challenges posed by different physiological states. Aiming for further performance enhancement, the classifier was trained using partial post-exercise data from subjects who were not part of the evaluation set. This training approach enabled the classifier to adapt to the characteristics of the evaluation set, enhancing its capability to recognize ECG signals from diverse states. Our proposed approach outperforms current methods, achieving superior accuracy in exercise recognition while maintaining excellent accuracy in resting states.

M.Sc. student under the supervision of Dr. Danny Lange and Dr. Kfir Y. Levy.