Research On Optimization Algorithm Based On Machine Learning In Intrusion Detection

Intrusion detection has received much attention since its introduction.In recent years,the wave of machine learning has brought fresh blood to the research of intrusion detection technology,and has been widely used.However,due to people's work and life are increasingly dependent on the Internet,the network information is intertwined and complicated,how to extracts useful information from massive data and analyzes it becomes the key to solving big data problems at present.The traditional feature selection method can improve the generalization ability of the algorithm to a certain extent by data dimensionality reduction,but it ignores the role of some important features in the classification,which leads to an imbalance between the intrusion detection rate and the false positive rate;For small categories of samples,most of the existing intrusion detection methods are not satisfactory in the performance of detection accuracy.Aiming at solving the above problems,this paper proposes a new feature selection method based on machine learning.The main research contents and innovations of the thesis are as follows:(1)Due to the diversification of sample feature attributes,the traditional methods can not represent the relationship between samples and feature attributes well,to solve such problems,this paper introduces the concept of hypergraph based on Fisher score feature selection,a two-level hybrid intrusion detection method based on FS-HG feature selection is proposed.In the training phase,the method combines the Fisher score and the Helly attribute of the hypergraph to feature selection,and obtains sample features that are more conducive to classification.In the test phase,random forests and improved K-means are used as joint classifiers,the aim is to avoid the influence of individual outliers on cluster center changes by adding density threshold,so that the clustering results reach higher accuracy.The experimental results show that the method of cascaded classifier can ensure that the algorithm achieves a good balance between detection rate and false positive rate,the detection accuracy of most categories is also higher than that of individual classifier.(2)In view of the long training time and easy over-fitting of traditional neural network models,an intrusion detection method based on asymmetric deep belief network is proposed.In the training phase,the method initializes the parameters in the ADBN model in an asymmetric manner,first,the parameters of the encoder in the ADBN model are obtained by training the DBN,and then initializes the parameters of the decoder through the values obey normal distribution and reconstruct the original data,finally,the parameters of the model are adjusted by calculating the reconstruction error.In the test phase,the data is extracted by using the encoder parameters in ADBN and used as the input data of the classifier to achieve the feature selection.The experimental results show that compared with the traditional DBN algorithm,the proposed method has greater advantages in detection rate and false positive rate,and it also achieves better detection accuracy for small class samples.