Biometric Authentication from ECG Signals on Wearable Devices

Abstract

Biometric authentication is currently being used for numerous devices; such as mobile phones, computers, etc. However, for now, the only authentication methods for wearable devices are those of passwords and pin codes. The newest instance of the Apple Watch series 4. has an integrated Electrocardiogram (ECG) recording possibility that could be used for biometric authentication. Having the possibility for biometric authentication on wearable devices could potentially provide seamless authentication applications as the wearable device is always on standby. The objective of this thesis was to test biometric authentication based on ECG signals recorded on wearable/mobile devices. By collecting data from a set of volunteers with recordings performed under different circumstances such as; resting heart rate, increased heart rate after exercise, and noisy signals while in motion. By performing denoising and feature extraction, various machine learning models were trained and evaluated to provide a classification model that performed well on the variety of ECG signals. The classification model was further used to present a biometric authentication system. The biometric authentication system presented in this thesis was tested on three different sets of acquired ECG data. Biometric authentication based on ECG signals recorded with resting heart rates correctly authenticated 17/19 subjects, resulting in an acceptance rate of 89.5%. For the recordings after physical activity and in motion, the authentication system correctly authenticated 52.6% (10/19) and 31.6% (6/19) of the subjects. An additional subject that had been excluded from the system did not get authenticated for either of the different recordings. Overall, no subjects were misclassified as other subjects.