Research On Anomaly Detection For Industrial Control Network Based On Improved OCSVM

In order to accomplish the detection and identification of outliers in industrial control networks,and address the problem that the abnormal threshold in the traditional network anomaly detection algorithm needs to be manually defined,and avoid the defect in the one-class support vector machine(OCSVM)that cannot be extended in the high-dimensional massive data.A deep learning anomaly detection algorithm(VAE-OCSVM)combined with Variable Auto-Encoder(VAE)and OCSVM is proposed in this article.In the anomaly detection process,the model firstly uses the VAE model to learn the characteristics of the data distribution,and encodes the input normal sample data to represent the distribution of the original normal sample data through low-dimensional coding.Secondly,the low-dimensional coding representation and the reconstruction error of the VAE model are merged into new inputs,increasing the feature information of the data points and improving the data reparability.Then,the traditional OCSVM's Hinge loss objective function and random Fourier feature are established to fit the RBF kernel,the OCSVM model is represented and solved by deep neural network and gradient descent method.Moreover,the fusion input is used as the input of the network,and the solution is obtained by solving Parameter information of the network model.Finally,the decision function of the OCSVM model is constructed by using the solved parameter information.When the function value of the detected object in the decision function is positive,the detection object is normal sample data;if the function value is negative,the value is outlier that different from the normal data.The experimental results of industrial control network security data show that the algorithm can effectively detect the outliers in the industrial control network data.In addition,compared with the mainstream anomaly detection algorithm,the VAE-OCSVM algorithm has better recognition ability than the traditional OCSVM anomaly detection algorithm.There are 22 figures,10 tables and 84 references in this paper.