Evaluation of data processing techniques for unobtrusive gait authentication

Abstract

The growth in smartphone usage has led to increased storage of sensitive data on these easily lost or stolen devices. In order to mitigate the effects of users who ignore, disable, or circumvent authentication measures like passwords, we evaluate a method employing gait as a source of identifying information. This research is based on previously reported methods with a goal of evaluating gait signal processing and classification techniques. This thesis evaluates the performance of four signal normalization techniques (raw signal, zero-scaled, gravity-rotated, and gravity rotated with zero-scaling). Additionally, we evaluate the effect of carrying position on classification. Data was captured from 23 subjects carrying the device in the front pocket, back pocket, and on the hip. Unlike previous research, we analyzed classifier performance on data collected from multiple positions and tested on each individual location, which would be necessary in a robust, deployable system. Our results indicate that restricting device position can achieve the best overall performance using zero-scaling with 6.13% total error rate (TER) on the XY-axis but with a high variance across different axes. Using data from all positions with gravity rotation can achieve 12.6% TER with a low statistical variance.