面向网络入侵检测数据的对抗样本生成方法

Adversarial examples generation method for network intrusion detection data

选用Deepfool以及JSMA(jacobian-based saliency map attack)算法,在攻击特征中加入不影响攻击特性的定向扰动,通过白盒攻击生成对抗样本。通过实现扰乱检测模型的判断,从而躲过特征检测,为入侵检测模型提升自身鲁棒性提供了更为丰富的训练样本。

This paper proposes to add directional perturbations having no impact on results to attack characteristics with Deepfool and JSMA algorithms. Adversarial samples are generated by white-box attacks so that they can interfere with the judgements of models to bypass feature detection. Our work provides intrusion detection models with more training samples. As a result, the robustness of intrusion detection models is improved.