Date of Award
Winter 2020
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Computational Analysis and Modeling
First Advisor
Vir Phoha
Abstract
Behavioral biometrics, such as keystroke dynamics, are characterized by relatively large variation in the input samples as compared to physiological biometrics such as fingerprints and iris. Recent advances in machine learning have resulted in behaviorbased pattern learning methods that obviate the effects of variation by mapping the variable behavior patterns to a unique identity with high accuracy. However, it has also exposed the learning systems to attacks that use updating mechanisms in learning by injecting imposter samples to deliberately drift the data to impostors’ patterns. Using the principles of adversarial drift, we develop a class of poisoning attacks, named Frog-Boiling attacks. The update samples are crafted with slow changes and random perturbations so that they can bypass the classifiers detection. Taking the case of keystroke dynamics which includes motoric and neurological learning, we demonstrate the success of our attack mechanism. We also present a detection mechanism for the frog-boiling attack that uses correlation between successive training samples to detect spurious input patterns. To measure the effect of adversarial drift in frog-boiling attack and the effectiveness of the proposed defense mechanism, we use traditional error rates such as FAR, FRR, and EER and the metric in terms of shifts in biometric menagerie.
Recommended Citation
Wang, Zibo, "" (2020). Dissertation. 844.
https://digitalcommons.latech.edu/dissertations/844