【作者】 王君； 吕露； 苏建忠； 王文海； 袁健全； 张泽银； 张正辉； 刘兴高；

【Author】 Jun Wang;Lu Lv;Jianzhong Su;Wenhai Wang;Jianquan Yuan;Zeyin Zhang;Zhenghui Zhang;Xinggao Liu;State Key Laboratory of Industrial Control Technology,Control Department,Zhejiang University;Tianjin Jinhang Institute of Technical Physics;Science and Technology on Complex System Control and Intelligent Agent Cooperative Laboratory;Mathematics Department,Zhejiang University;

【机构】 浙江大学NGICS大平台浙江大学控制学院工业控制技术国家重点实验室； 天津津航技术物理研究所； 复杂系统控制与智能协同技术重点实验室； 浙江大学数学科学学院；

【摘要】 作为工控系统安全的重要子领域,入侵检测一直是国际上的难点和前沿,其关键在于如何高精度地识别出不同的网络入侵行为。本文提出了一种新颖的深度度量学习入侵检测方法,与传统方法相比,能够有效地提取特征,并通过计算各类损失,从而形成紧凑的聚类空间,便于识别出各种不同的网络入侵行为。以当前主流的KDD99数据集为例进行了实验,在Normal,DoS,PRB,U2R,R2L等5个类别上分别达到了99.34%,100%,93.18%,92.86%,88.06%的识别精确率,与当前主流算法相比,后4个类别的识别都达到了最高的精确率,表明了所提出方法的有效性.更多还原

【Abstract】 As an important sub-field of industrial control system security,intrusion detection has always been an international difficulty and frontier,the key to which is how to identify different network intrusions with high precision.Compared with traditional methods,deep metric learning can effectively extract features,and by calculating various losses,a compact clustering space is formed to facilitate the identification of various network intrusions.The current mainstream KDD99 dataset is used as to conduct the experiment,and the precision rate of Normal,Dos,PRB,U2 R,R2 L achieve 99.34%,100%,93.18%,92.86%,88.06%,respectively.Compared with other literature results,the precision of the last 4 categories achieve the highest accuracy rate,which validate the effectiveness of the proposed approach.更多还原