An autonomous decision-making framework for gait recognition systems against adversarial attack using reinforcement learning

Abstract

Gait identification based on Deep Learning (DL) techniques has recently emerged as biometric technology for surveillance. We leveraged the vulnerabilities and decision-making abilities of the DL model in gait-based autonomous surveillance systems when attackers have no access to underlying model gradients/structures using a patch-based black-box adversarial attack with Reinforcement Learning (RL). These automated surveillance systems are secured, blocking the attacker's access. Therefore, the attack can be conducted in an RL framework where the agent's goal is determining the optimal image location, causing the model to perform incorrectly when perturbed with random pixels. Furthermore, the proposed adversarial attack presents encouraging results (maximum success rate = 77.59%). Researchers should explore system resilience scenarios (e.g., when attackers have no system access) before using these models in surveillance applications.

Keywords: Adversarial attack; Autonomous surveillance system; Decision-making ability; Reinforcement learning.