In black-box scenarios, the adversarial attack algorithms based on iterative optimization perform best. But it may give a false sense of model robustness due to the design of inefficient queries. The adversarial attack algorithm based on multi-objective evolution optimization has been proven to be very effective for the low-dimensional images. However, when the attack space dramatically increases for the high-dimensional color images, the evolutionary efficiency is limited, and it needs more inefficient queries to generate adversarial examples. In this paper, we propose an efficient black-box adversarial attack approach for high dimensional images based on multi-objective optimization (MOO-HD), which includes some novel strategies to solve the above problems. We also propose the strategy of “The transformation of the pixel block with a random step size” to reduce the attack space. The experimental results on three image datasets with different dimensions show that our algorithm can achieve a higher success rate with fewer queries.

在黑盒场景中，基于迭代优化的对抗性攻击算法表现最好。但是由于低效查询的设计，它可能会给模型健壮性带来错误的感觉。基于多目标进化优化的对抗性攻击算法已被证明对低维图像非常有效。然而，当高维彩色图像的攻击空间急剧增加时，进化效率有限，需要更多低效的查询来生成对抗样本。在本文中，我们提出了一种基于多目标优化（MOO-HD）的高维图像有效黑盒对抗攻击方法，其中包括一些解决上述问题的新策略。我们还提出了“以随机步长变换像素块”的策略来减少攻击空间。在三个不同维度的图像数据集上的实验结果表明，我们的算法可以用更少的查询获得更高的成功率。