Light Attack: A Physical World Real-Time Attack Against Object Classifiers

Abstract

It is well known that deep neural networks (DNNs) are vulnerable to adversarial examples. In the digital world, most of the existing work makes classifiers or detectors fail by adding perturbations that are imperceptible to humans. In the physical world, existing work mostly invalidates classifiers or detectors by adding large unrealistic perturbations. In this paper, we attack the target classifiers from a new perspective of light attack. First, we define three lighting modes (Line, Point, Area), then we control the wavelength, intensity and position of light to accomplish adversarial attack, and finally we conduct ablation experiments and compare our method with existing mainstream methods. Experiments demonstrate that our proposed method is effective in both digital and physical worlds, and in darker environments, our attack method is much better than existing ones. Light attack enriches the current family of the adversarial attack while paving the way for future defense against light attacks as well as the study of light attack laws in nature.