Application Of Machine Learning Algorithms In Cyber Physical System Security

Cyber-Physical System combines environmental awareness,embedded computing and network communications to form a multi-dimensional heterogeneous complex system which integrates real-time sensing,dynamic control and information services.Due to the high integration of information systems and physical systems,the entire system has become more flexible and open.A large number of node devices access to the information platform.While bringing convenience,the environment is gradually becoming more complicated.The scenario of CPS being attacked is diversified,and the types of attacks are developing in a highly destructive and intelligent manner.Therefore,research on CPS security issues is becoming increasingly important.Based on the existing research results,this thesis comprehensively considers the requirements of CPS security,conducts research on CPS attack detection and defense.This thesis mainly completed the following tasks:(1)From the perspective of the attack results,three attack models are established.The Tennessee-Eastman process is used as the experimental platform,and the established attack model is used in the experimental platform to generate normal data and attacked data to provide basis for attack detection and data prediction.(2)According to the characteristics of CPS data,this thesis proposed a method based on Sparse Autoencoder(SAE)to reduce the dimension of CPS.New feature representation was reconstructed by using an unsupervised method.Support Vector Machine(SVM)as a detector to detect attacks.At the same time,an improved bacterial foraging algorithm is proposed to optimize the parameters of the support vector machine in order to set the kernel function parameters and penalty factors.Finally,the Tennessee-Eastman process data set and the gas pipeline data set are used to verify the algorithm proposed in this thesis.The results show that the proposed method has good performance in attack detection.Compared with K-nearest Neighbor(KNN)algorithm,logistic regression(LR)algorithm,and SVM algorithm,the proposed method has higher performance.(3)Consider that the attack will make the data transmitted by the CPS unreliable or undeliverable,which will cause the actuator to malfunction.A method of improving the Long Short-Term Memory(LSTM)network is proposed to predict the control data and provide reliable data for the actuator after the attack.First,the input gate and forget gate of the LSTM network were combined to simplify the weight parameters and improve the structure of the network.Then the CPS data prediction process is designed and applied to Tennessee-Eastman process,and compared with the prediction performance of LSTM network and Gated Recurrent Unit(GRU)network.The results show that the algorithm proposed in this thesis is more accurate.