Cyber-physical Attacks Detection And Location Of Water Distribution Systems Based On Deep Learning

With the rapid development of intelligent informatization,more and more traditional key infrastructure is upgrading to a cyber-physical system（CPS）,including power systems,communication systems,and water distribution systems（WDS）.In particular,WDS plays an extremely important role in the production and life of the national economy,the defense industry and the people.Since WDS’s intelligence is getting higher and higher,the risk of cyber-physical attack（CPA）is getting bigger and bigger.Water supply network is affected by CPA has become an important issue in the world’s countries.The consequences will not only lead to the waste of water resources and the loss of economic property,and even have a serious impact on people’s normal life,national security.Therefore,ensure the safety of the water supply pipe network system,timely detection of the CPA received by the water supply network and positioning the attacked components has become an important need.This paper proposes the detection and location algorithm based on the cyber-physical attacks in the water distribution systems based on stacked-autoencoder,using SAE to build a depth learning network model.This thesis uses SAE to take the form of high-dimensional data characteristics in the form of pre-training by SAE.This thesis uses historical observation data for the normal hydraulic operation process of the water distribution system to train SAE model.In the training completed model input exception data,the reconstructed data will have a large deviation value.In addition,this thesis uses a moving average（MA）algorithm based on the statistical window to detect the error between the original input data and the predicted sensor data,in order to identify and report an abnormality.The attack location of the corresponding components can be performed by analyzing the error.In the experiment,the results showed that SAE-based detection and positioning algorithms proposed in this thesis are superior to other methods,and can locate physical components affected by the attack.Finally,the SAE detection and location algorithm are discussed in the statistical window size and the detection performance of the network structure complexity.Considering that the SAE attack detection and the location algorithm are mapped to unknown hidden vector spaces when extracting data features,which affects the extraction of data.Depending on the theory of generating adversarial network（GAN）related theory,GAN can constrain the feature space of the input data to the Gaussian vector space,so this thesis proposes GAN attack detection and location algorithm.And considering that the ordinary GAN model is only used to generate data,and in this thesis,it is based on data reconstruction to perform anomaly detection.The algorithm performs extraction of data characteristics by mapping training samples to Gaussian vector space.The experimental results show that the results show that the algorithm has strong detection performance,and can be located to the physical components of the attack-based target network.