A Meta-Heuristic Optimization Based Less Imperceptible Adversarial Attack on Gait Based Surveillance Systems

Abstract

In today's digital realm, global safety concerns have given rise to intelligent surveillance technologies. Currently, Gait-based surveillance systems have lately got increasing attention and are widely used because they employed human distinctive and behavioral characteristics as well as recognize them without their cooperation. They play a vital role in smart video monitoring systems and have a broad array of applications, especially in public security applications. Gait recognition is the process of validating a person looking at the way they walk. Behind these surveillance systems, superior computer vision-based deep learning algorithms are deployed for the effective identification of individuals. On the other hand, the vulnerable nature of these algorithms is the main factor for potential security threats in these surveillance systems. Different researchers design several kinds of adversarial attacks for different computer vision domains to exploit their vulnerabilities. In most of these attacks, the resulting adversarial noise in form of either patches or pixels is apparent and clearly identifiable by the human naked eye. From this line of research, this study presents a new variant of the attack, to highlight the vulnerability of gait recognition systems. The adversarial noise in the form of one pixel is added to that location of Gait energy images (GEI) which are less imperceptible by humans. More specifically, this can be accomplished using the Grey wolf optimization (GWO) method, which takes a GEI image and a calculated perturbation as input and starts working to find the optimum location on a given GEI image. This optimum candidate/location is the one where the texture is not excessively affected when it is disrupted by computed perturbation, and as a result, the attack is less perceptible. Moreover, the study investigates that CNN's based gait recognition systems encountered severe potential threats even when the noise is less imperceptible by the human naked eye.