Exploiting vulnerability of convolutional neural network-based gait recognition system

In today’s era of advanced technologies, the concerns related to global security have led to video surveillance gadgets. Human gait recognition as a biometric is considered an evolving technology for intelligent surveillance monitoring. This research study exploits vulnerabilities associated with a convolutional neural network (CNN)-based gait recognition system under various walking conditions involving clothing, carrying items, and speed. In the first stage, we design a CNN model capable of identifying individuals based on their gait characteristics. Subsequently, in the next stage, we design a five-pixel adversarial attack in which we perturb the gait features of individuals computed using the fast gradient sign method. The resulting perturbation is added to only five random pixels to create naturalistic adversarial samples similar to the original samples. Further, the main aim of this study is to determine and analyze the performance of the CNN-based gait recognition system under an adversarial attack. The research concludes that gait recognition systems based on CNN models are highly susceptible to adversarial attacks, motivating researchers to design defense mechanisms to mitigate the effect of these attacks.