A Robust Authentication Algorithm for ECG Wavelet with Varying Characteristics

Abstract

Human authentication based on electrocardiogram (ECG) has been a remarkable issue for recent twenty years. Several anatomical and physiological patterns, such as fingerprints, face, and iris, have been introduced and applied already in some smart devices. However, those techniques are suffering with the forgery problems. Recently, the new generation of biometric authentication modality, which are biomedical signals that typically used for clinical diagnostic purposes, has been suggested for these problems. Among the biomedical signals, ECG has been the most robust modality for biometric authentication, because it is recorded in non-invasive, simple, effective, and low-cost procedure. In this study, infinite feature selection was used to decide discriminative features for verification, and the Mahalanobis distance, which considers both the mean difference and standard deviation of the distribution, was used to verify the test data is from genuine or imposter. Verification performance was evaluated by the measure of equal error rate (EER), and compared with state-of-the-art authentication systems. ECG data was acquired from 105 healthy subjects in three modulated situations

the subjects were asked to record their ECG in two different day (temporal difference), in two different posture (postural difference), and in pre-exercise, post-exercise (physical activity status difference) situations. The signals have been obtained for about 5 minutes in each acquisition status. Both enrollment and test data was pre-processed before the verification step is conducted. Pre-processing steps are divided into three sub-steps: Filtering, Heartbeat Extraction, Outlier Removal. After pre-processing, best features were selected using Infinite Feature Selection in the enrolled set. The order of selected features was enrolled in the system for the user and it was re-used in verification step when the test set was acquired. A pre-processed heartbeat of test set is reorganized in system-enrolled order and top M features are selected. These test features are compared with the enrolled features by the Mahalanobis distance, and if the distance is smaller than enrolled threshold the system claims that the test user is genuine. In this study, only 1-to-1 verification was evaluated by comparing EER of other state-of-the-art works. In 1-to-1 verification, identity has already been declared among the N identities, and the system verifies whether the test user matches with the identity. When we averaged over all participants, we obtained an EER of 1% in the time-varying situation, 2.21% in the posture change situation, and 5.80% after the exercise. This results is better than the result from state-of-the-art techniques, and especially after exercise, EER was about 12 % lower than the best algorithm among the previous works.