Fooling A Deep-Learning Based Gait Behavioral Biometric System

摘要: We leverage deep learning algorithms on various user behavioral information gathered from end-user devices to classify a subject of interest. In spite the ability these techniques counter spoofing threats, they are vulnerable adversarial attacks, where an attacker adds noise input samples fool classifier into false acceptance. Recently, handful mature like Fast Gradient Sign Method (FGSM) have been proposed aid white-box has complete knowledge machine model. On contrary, we exploit black-box attack biometric system based gait patterns, by using FGSM and training shadow model that mimics target system. The limited no real being authenticated, but induces acceptance in authentication. Our goal is understand feasibility what extent models would contribute its success. results manifest performance highly depends quality model, which turn impacted key factors including number queries allowed order train experimentation revealed strong relationships between performance, as well effect iterations used create instance. These insights also shed light deep-learning algorithms' shareability can be exploited launch successful attack.