Publication: Generalized electrocardiogram biometrics based on encoder representations from transformer with augmented dataset generation
Loading...
Date
2024-10-01
Authors
Chee, Kai Jye
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Electrocardiogram (ECG) biometrics enhances security by complementing
other identity proving methods, due to the inherent difficulty of falsifying an internal
signal. Traditional machine learning approaches have various limitations, including
issues like inadequate training data and overlooking the dependencies between
enrolment and query segments. Furthermore, conventional identification classifiers
face a trade-off between accuracy and adaptability to enrolment changes, often
requiring retraining upon new enrolments. This research introduces Multi-database
Training Examples Generation (MTEG) to generate abundant training datasets from
multiple ECG databases. This research also presents the Enrolled-Query ECG Pair
Feature Extractor (EQFE) to extract inter-segment dependencies of the enrolment and
query segments, and Self-Attention Identification Classifier (SAIC) that considers all
enrolment subjects in the identification task, while also accommodating changes to the
enrolment subjects without necessitating model retraining. The incorporation of ten
databases in MTEG improves authentication accuracy from 71.54% to 99.57% in the
PTB Diagnostic ECG Database, compared to using a single database. Furthermore, the
presence of EQFE enhances authentication accuracy from 70.94% to 98.12% in the
Stress Test Database, relative to its absence. Additionally, the utilization of SAIC
results in a notable improvement in identification accuracy, rising from 92.87% to
96.29% in the Atrial Fibrillation Database, compared to its absence.