摘要： 为高效地寻找基于决策的黑盒攻击下的对抗样本，提出一种利用模型之间的迁移性提升对抗起点的方法。通过模型之间的迁移性来循环叠加干扰图像，生成初始样本作为新的攻击起点进行边界攻击，实现基于决策的无目标黑盒对抗攻击和有目标黑盒对抗攻击。实验结果表明，无目标攻击节省了23%的查询次数，有目标攻击节省了17%的查询次数，且整个黑盒攻击算法所需时间低于原边界攻击算法所耗费的时间。

关键词: 黑盒攻击, 对抗样本, 迁移性, 初始样本, 边界攻击, 无目标攻击, 有目标攻击

Abstract: In order to efficiently find the adversarial samples under the decision-based black box attacks, a method using the mobility between models is proposed to enhance the adversarial starting point. The mobility is used to circularly superimpose interference images, and samples are generated as a new starting point for boundary attacks. Thus the decision making-based non-target adversarial black box attacks and targeted adversarial black box attacks are realized. Experimental results show that the query times required for the non-target attacks is reduced by 23%, and that required for the targeted attacks is reduced by 17%. Moreover, the whole black box attack algorithm takes less time than the original boundary attack algorithm.

Key words: black box attack, adversarial sample, mobility, initial sample, boundary attack, non-target attack, targeted attack

中图分类号: 

• TP183