Adversarial Examples with Missing Perturbation Using Laser Irradiation

There exist Adversarial Examples (AEs) that intentionally misclassified images of neural networks to a different class. We propose a method of attack by AEs that uses lasers to facilitate control of misclassification. A previous work has shown that road signs can be misclassified by irradiating them with lasers. We focus on the effective attack in the situations that the attacker can prepare the road signs. Namely, in our attack, the attacker prepares the road signs made by using AEs with missing perturbation, and makes misclassified by using laser irradiation in effective. In this paper, we present principle and method for misclassifying the road signs made by the AEs with missing perturbation by using laser irradiation. In addition, we demonstrate the effectiveness of the proposed method by experiments.