| With the introduction of the concepts of Industrial Internet,Industry 4.0,and China Manufacturing 2025,the industrial production process will be more intelligent and efficient,and the control of the production process will be more complicated and open,but it increases the safety risks of industrial production at the same time.As a high-value attack target,attacks in industrial control systems have become more and more frequent in recent years,so it is imperative to study the safety of the industrial control systems.Attacks can be detected with the intrusion detection system specifically for industrial control systems.With the development of machine learning,more and more algorithms have been applied to intrusion detection.These higher-precision algorithms have their vulnerability although they can get better results in intrusion detection.Therefore,to design an algorithm with excellent detection effect and consider the robustness of this algorithm,three aspects of research work have been carried out.First,a method for intrusion detection using the recurrent neural network is proposed.Construct the recurrent neural network model and the way of data input and output,training the model with the normal data which is generated in the industrial production process.And obtain the prediction result of the recurrent neural network,then get the intrusion state of the system by calculating the distance between the data in reality and the predicted outcome.Then,Commence the vulnerability analysis of the intrusion detection model and defined adversarial examples in industrial control network intrusion detection.We are proposing an adversarial examples generation method for the regression problem.The disturbance is added to the original examples continuously through the idea of iteration,and guide the orientation of the addition with the Jacobian matrix.Finally,analyze the characteristics of the attack which using the vulnerability of the model,two methods for defending adversarial examples are proposed.One is using the features of the adversarial examples generation targeted,collect new data persistently to update the parameters of the neural network.The other method is using the characteristics of the difference between the adversarial examples and normal examples.Learn the characteristics of the normal sample with the autoencoder,and combines the recurrent neural network model for intrusion detection.This paper implements the intrusion detection algorithm proposed by it and experiments on the complex industrial control system dataset,to verify the effectiveness of the proposed intrusion detection algorithm.As for the adversarial examples under the regression problem,we generate effective adversarial examples on the data set.And verify two defense methods proposed in this paper.Extract the subset from the data set to verify the effectiveness of the defense method for updating the neural network parameters.Autoencoder's defense effect is verified using the generated adversarial examples. |
