Black-box l1 and £2 Adversarial Attack Based on Genetic Algorithm

It has been shown that multimodal text-vision large language models (LLMs) are vulnerable to visual adversarial examples. An adversarial example is a delicately perturbed input intended to cause deep neural networks to behave incorrectly. Currently, adversarial attacks can be categorized into two types based on the knowledge required to perform an attack: black box and white box. Most existing black-box attacks only try to generate adversarial examples within a given
$\ell_{2}$
or
$\ell_{\infty}$
pertur-bation budget. We propose a novel black-box attack to address this research gap to generate
$\ell_{1}, \ell_{2}$
adversarial examples with minimal perturbation. To this end, our attack solves a constrained optimization problem by converting it to an unconstrained one and then finds its solution using the genetic algorithm. Extensive experiments on three test benches have been conducted to evaluate its performance. To facilitate reproducing the results presented in this work, we make the code for the experiments publicly available at https://github.com/sjysjy1/MPBA.