摘要
文中提出了一种提高深度学习模型对抗样本攻击鲁棒性的网络入侵检测方法。对比了4种不同对抗样本生成技术在两种不同类型攻击上的表现,从归一化的特征空间来评估网络的安全性。将传统的手工固定阈值进行回归模型学习,通过后处理变换转变为自适应阈值。利用弹性网络方法进行对抗样本生成和网络入侵检测优化,在尽可能小的输入扰动下实现混淆入侵检测系统的分类,增强鲁棒性。
A network intrusion detection method is proposed to improve the robustness of deep learning model against sample attacks.The performance of four different counter sample generation techniques on two different types of attacks is compared,and the network security is evaluated from the normalized feature space.The traditional manual fixed threshold method for regression model learning is transformed into adaptive threshold classifier by post-processing transformation.The resilient network is used to generate the counterwork samples and optimize the network intrusion detection network to achieve the classification ability and robustness of the intrusion detection system with as little input disturbance as possible.
作者
胡声秋
李友国
高渊
吴玲丽
HU Shengqiu;LI Youguo;GAO Yuan;WU Lingli(China Mobile Chongqing Co.,Ltd.,Chongqing 401121,China)
出处
《电子设计工程》
2022年第11期50-54,59,共6页
Electronic Design Engineering
关键词
物联网安全
网络入侵检测
对抗样本生成
自适应阈值
IoT security
network intrusion detection
adversarial sample generation
adaptive threshold