基于CNN⁃CBAM的虚假数据注入攻击辨识研究

Research on False Data Injection Attack Identification based on CNN⁃CBAM

【目的】针对当前智能电网受到网络攻击时,未能及时锁定网络攻击位置实现快速部署防御策略的问题。【方法】文章提出了一种融合卷积注意力模块(CBAM)的卷积神经网络(CNN)模型(即CNN⁃CBAM模型)对虚假数据注入攻击(FDIA)位置进行检测。将FDIA的攻击辨识问题建模为一种多标签分类问题,CNN用于提取数据的空间特征,CBAM直接融合到CNN模块的卷积操作后,不仅能从空间域的角度关注重要的参数信息,同时还考虑了通道域上的特征关系,从两个维度对输入数据进行注意力分配,以提升模型的性能。【结果】在电气和电子工程师协会(Institute of Electrical and Electronics Engineers,IEEE)14和IEEE118节点系统上对所提出的CNN⁃CBAM模型FDIA位置检测的性能进行验证,实验结果表明,CNN⁃CBAM在IEEE14和IEEE118节点系统上的FDIA位置检测率分别为98.25%和96.72%。【结论】与其他方法相比,文章所提出的CNN⁃CBAM模型能够有效地提取数据间的时空特性,提高FDIA存在性和攻击位置辨识精度,并具有更好的鲁棒性。

【Objective】It is always difficult to timely locate the location of the network attack and achieve rapid deployment of defense strategies when the smart grid is attacked by the network.【Methods】In order to solve this problem,this article proposes a Convolutional Neural Network(CNN)model that integrates Convolutional Block Attention Modules(CBAM)(CNN⁃CBAM)to detect False Data Injection Attack(FDIA)positions.The attack identification problem of FDIA is modeled as a multi label classification problem,where CNN is used to extract spatial features of the data.The CBAM module can be directly integrated into the convolution operation of the CNN module,which not only focuses on important parameter information from the perspective of spatial domain,but also considers feature relationships in the channel domain,and allocates attention to the input data from two dimensions to improve the performance of the model.【Results】The performance of the proposed CNN⁃CBAM network FDIA position detection model is verified on Institute of Electrical and Electronics Engineers(IEEE)14 and IEEE118 node systems.The experimental results show that the FDIA position detection rates of CNN⁃CBAM on IEEE14 and IEEE118 node systems are 98.25%and 96.72%,respectively.【Conclusion】Compared with other methods,the CNN⁃CBAM network model proposed in this paper can effectively extract the spatiotemporal characteristics between data,with improved existence of FDIA.It also improves the accuracy of attack location identification with better robustness.