Institutional Scholarship

On the Security of Wearable Sensor-based Gait Authentication

Show simple item record

dc.contributor.advisor Kumar, Rajesh Mo, Jun 2021-07-12T11:57:41Z 2021-07-12T11:57:41Z 2021
dc.description.abstract Gait is the style or manner in which someone walks. The usage of gait analysis could lead to the next breakthrough in biometric authentication systems. Gait is an ideal biometric for authentication systems for non-intrusive and naturally produced unique personal identifiers. The study of gait has come a long way since the first video-based analysis in the 1990s. Recently, sensor-based gait authentication systems have become popularized as smartphones and smartwatches have become more accessible to the average consumer. While with great potential, gait-based authentication studies still require much analysis. Many researchers have explored potential methods that exploit gait instability as a modality to create attack methods to circumvent machine-learning-based authentication models. The literature review examines three different attack scenarios that could circumvent a gait-based security system and a defense method to mitigate said attacks. One attack proposes a K-means-based algorithm that extracts similar gait cycles between users. Another has shown that if an attacker knows the vector space of an authentication model, it can use uniform random inputs to find an accepting sample even if it has a low false-positive rate. The last attack proposes a treadmill-assisted imitation attack that helps an impostor mimic aspects of the target's movements. One group of researchers offer a defense method that employs multiple devices (smartwatch and smartphone) that strengthens the robustness of zero-interaction authentication systems. For my proposed work, I explore using a generative model to reinforce the robustness of a gait-based authentication system. In response to the random input attacks, I propose a defense method using CTGAN to produce synthetic data to augment the original dataset. We augment the original dataset in three different ways: (1) Supplementing the genuine user with its own synthetic data, (2) Use the genuine user's data for the impostor class, and (3) supplement both the genuine and impostor class with synthetic data based on its respective class. We prove that CTGAN generated data works better to protect against random input attacks than just beta-distribution noise. This finding applies to not just data collected from smartphones but also to smartwatch data.
dc.description.sponsorship Haverford College. Department of Computer Science
dc.language.iso eng
dc.title On the Security of Wearable Sensor-based Gait Authentication
dc.type Thesis
dc.rights.access Dark Archive until 2022-01-01, afterwards Open Access.

Files in this item

This item appears in the following Collection(s)

Show simple item record Except where otherwise noted, this item's license is described as



My Account