| In recent years,the industrial control industry is gradually moving towards the direction of digital and intelligent development.While the industrial control system is open and interconnected,it is also facing more complex security challenges.The research of intrusion anomaly detection in industrial control system faces the problem of too few underlying business data sets,especially the data sets containing effective underlying business negative samples.With the introduction of deep learning into industrial control system,the edge computing platform commonly used in industrial control system often can not meet the requirements of deep learning model for storage and computing resources.In this paper,the negative sample expansion algorithm and deep neural network pruning algorithm of industrial control system are studied.The innovations and research results are as follows:(1)A generative countermeasure network based on double discrimination model is proposed.Aiming at the underlying business data of industrial control system,an efficient and high-quality attack sample generation method is established.Different from the traditional generative countermeasure network,this paper adds an intrusion discrimination model in the discrimination link,which not only ensures that the generated attack samples are similar to the seed attack samples,but also ensures the intrusion of the attack samples.After verification,compared with the samples generated by the traditional generative countermeasure network and unbalanced samples,the samples generated by this method have higher similarity rate with the original samples,stronger authenticity,and can effectively reduce the false alarm rate and improve the accuracy of the anomaly detection model.(2)A pruning algorithm based on Poisson distribution is proposed.In the iterative process of the deep neural network,based on the Poisson distribution hypothesis,the historical changes of parameters in the convolution template are calculated by the Poisson probability formula.In the training process,the template with less parameter changes is deleted as soon as possible,which can reduce the size of the model and accelerate the training process.After verification,the pruning rate of this method in the industrial control data set can reach 95%,the accuracy loss rate of Top1 is only 1.06%,and the pruned model is only 3.7% of the original model.Compared with other algorithms on the image data set,this paper retains higher accuracy under the same pruning rate,and the retraining process is faster.So as to meet the security and performance requirements of edge cloud collaboration in industrial control scenarios. |
