摘要： 为提高分类器在对抗性环境和训练阶段的抗攻击性,提出一种新的攻击模拟算法。通过拟合成员分类器模拟并获取最差情况攻击使用的决策边界,根据阈值设定去除性能较差的成员分类器,使最终攻击结果优于模仿攻击算法。实验结果表明,该算法无需获取目标分类器的具体信息,在保证分类准确率的同时具有较高的安全性。

关键词: 分类器, 对抗性环境, 攻击模拟算法, 最差情况攻击, 模仿攻击

Abstract: In order to improve the anti-aggressive capability of classifiers in the adversarial environment and training stage,this paper proposes a new attack simulation algorithm.The member classifiers are fitted to simulate and get the decision boundary used by the worst case attack,and the member classifier with poor performance is removed according to threshold setting.The final result of the proposed algorithm is superior to that of the mimicry attack algorithm.Experimantal result shows that this algorithm has no need to get the specific information of the target classifier,and it has higher security while maintaining the accuracy of classification.

Key words: classifier, adversarial environment, attack simulation algorithm, worst case attack, mimicry attack

中图分类号: 

• TP18