Back-Propagation-Free Adversarial Attacks with Zero-th Order Gradient Estimation

Adversarial attack is an indispensable way to evaluate and improve machine learning model robustness. Most attacks use the first-order gradient methods, which rely on back-propagation that can be computationally expensive. In this paper, we examine a zeroth-order gradient estimation method for back-propagation-free adversarial attacks. The paper evaluates the effectiveness of zeroth order gradient estimation using a variety of parameters and models. The paper serves as a proof of concept for zeroth order attacks and motivates further research to unveil the true potential of zeroth order adversarial attacks in a wide range of applications.