Abstract

Smartwatches, which contain an accelerometer and gyroscope, have recently been used to implement gait and gesture- based biometrics; however, the prior studies have long-established drawbacks. For example, data for both training and evaluation was captured from single sessions (which is not realistic and can lead to overly optimistic performance results), and in cases when the multi-day scenario was considered, the evaluation was often either done improperly or the results are very poor (i.e., greater than 20% of EER). Moreover, limited activities were considered (i.e., gait or gestures), and data captured within a controlled environment which tends to be far less realistic for real world applications. Therefore, this study remedies these past problems by training and evaluating the smartwatch-based biometric system on data from different days, using large dataset that involved the participation of 60 users, and considering different activities (i.e., normal walking (NW), fast walking (FW), typing on a PC keyboard (TypePC), playing mobile game (GameM), and texting on mobile (TypeM)). Unlike the prior art that focussed on simply laboratory controlled data, a more realistic dataset, which was captured within un-constrained environment, is used to evaluate the performance of the proposed system. Two principal experiments were carried out focusing upon constrained and un-constrained environments. The first experiment included a comprehensive analysis of the aforementioned activities and tested under two different scenarios (i.e., same and cross day). By using all the extracted features (i.e., 88 features) and the same day evaluation, EERs of the acceleration readings were 0.15%, 0.31%, 1.43%, 1.52%, and 1.33% for the NW, FW, TypeM, TypePC, and GameM respectively. The EERs were increased to 0.93%, 3.90%, 5.69%, 6.02%, and 5.61% when the cross-day data was utilized. For comparison, a more selective set of features was used and significantly maximize the system performance under the cross day scenario, at best EERs of 0.29%, 1.31%, 2.66%, 3.83%, and 2.3% for the aforementioned activities respectively. A realistic methodology was used in the second experiment by using data collected within unconstrained environment. A light activity detection approach was developed to divide the raw signals into gait (i.e., NW and FW) and stationary activities. Competitive results were reported with EERs of 0.60%, 0% and 3.37% for the NW, FW, and stationary activities respectively. The findings suggest that the nature of the signals captured are sufficiently discriminative to be useful in performing transparent and continuous user authentication.

Document Type

Thesis

Publication Date

2020-01-01

DOI

10.24382/758

Download

Share

COinS