【作者】 李超；

【导师】 杜晔；

【作者基本信息】 北京交通大学 ， 网络与信息安全（专业学位）， 2024， 硕士

【摘要】 传统的配电网络正逐渐演变成高度交织的信息物理系统,从而实现终端用户和发电公用事业之间的双向通信。然而,网络和物理系统之间的深度耦合也使配电网更容易受到外部网络安全威胁。作为一种典型的恶意攻击形式,虚假数据注入攻击(False Data Injection Attacks,FDIA)具有极高的隐蔽性和极大的威胁性,通过避开配电系统状态估计中的不良数据检测机制,破坏量测数据的完整性,导致有偏差的状态估计,最终干扰配电系统的正常运转,并造成巨大的经济损失。因此,解决此类攻击造成的网络安全问题至关重要。本文以配电网系统为研究对象,分别提出了虚假数据注入攻击的构造和检测方法,具体工作如下:(1)分析了配电网状态估计的实现机制,面向应用不良数据检测机制和深度神经网络检测方法的配电网系统,提出了一种最小成本对抗性FDIA(MCSAFDIA)构造方法。基于两阶段优化过程构建最小攻击向量,并针对最优目标节点实施攻击;设计了基于对抗攻击的扰动向量,对节点的状态偏移量进行攻击,以逃避不良数据检测机制和深度神经网络方法的双重检测。实验结果表明,该方法能够有效地绕过检测,且攻击成功率达到98.4%。(2)针对配电网虚假数据注入攻击量测实例少、数据不平衡等问题,提出了一种基于CGAN的数据增强模型。面向生成对抗网络存在训练过于自由而不可控、模式崩溃等情况,设计攻击类别标签以优化数据生成过程,进而提高数据生成的质量。基于CGAN模型学习量测样本的分布特征,并生成与真实量测数据无法区分的假量测样本(攻击样本)。实验结果表明,该模型生成的数据能够以高于99%的概率通过不良数据检测。(3)为提高FDIA的检测性能,提出了一种基于数据增强和CNN-Bi GRUSA的虚假数据注入攻击检测方法。首先使用上文基于CGAN的数据增强模型生成所需数据,从而平衡检测所用数据。其次设计了一种基于三层网络架构的检测模型CNN-Bi GRU-SA,提取预测特征与数据在高维空间的联系,构造时序序列的高维特征向量,并融合自注意力机制以强化对特征中重要信息的聚焦能力。实验结果表明,该方法的检测准确率达到98.8%,性能得到明显提升。更多还原

【Abstract】 Traditional power distribution networks are evolving into highly intertwined cyberphysical systems,enabling two-way communications between end users and generating utilities.However,the deep coupling between cyber and physical systems also makes distribution networks more vulnerable to external cybersecurity threats.As a typical form of malicious attacks,False Data Injection Attacks(FDIA)are highly concealed and extremely threatening.They destroy measurement data by avoiding the bad data detection mechanism in power distribution system state estimation.The integrity of the power distribution system leads to biased state estimation,ultimately interfering with the normal operation of the power distribution system and causing huge economic losses.Therefore,it is crucial to address the cybersecurity issues caused by such attacks.This thesis takes the distribution network system as the research object and proposes the construction and detection methods of false data injection attacks.The specific work is as follows:(1)The implementation mechanism of distribution network state estimation is analyzed,and a minimum cost adversarial FDIA(MC-SAFDIA)construction method is proposed for distribution network systems that apply bad data detection mechanisms and deep neural network detection methods.A minimum attack vector is constructed based on a two-stage optimization process,and the attack is carried out against the optimal target node;a perturbation vector based on adversarial attacks is designed to attack the state offset of the node to evade the bad data detection mechanism and deep neural network method.Double detection.Experimental results show that this method can effectively bypass detection,and the attack success rate reaches 98.4%.(2)Aiming at the problems of few false data injection attack measurement examples and data imbalance in distribution network,a data enhancement model based on CGAN is proposed.Faced with situations such as too free and uncontrollable training and mode collapse in generative adversarial networks,attack category labels are designed to optimize the data generation process and thereby improve the quality of data generation.Based on the CGAN model,it learns the distribution characteristics of measurement samples and generates fake measurement samples(attack samples)that are indistinguishable from real measurement data.Experimental results show that the data generated by this model can pass bad data detection with a probability higher than(3)In order to improve the detection performance of FDIA,a false data injection attack detection method based on data enhancement and CNN-Bi GRU-SA is proposed.First,use the above CGAN-based data augmentation model to generate the required data to balance the data used for detection.Secondly,a detection model CNN-Bi GRUSA based on a three-layer network architecture is designed to extract the connection between prediction features and data in high-dimensional space,construct a highdimensional feature vector of the time series sequence,and integrate the self-attention mechanism to enhance the feature recognition The ability to focus on important information.Experimental results show that the detection accuracy of this method reaches 98.8%,and the performance is significantly improved.更多还原