Design And Implementation Of Network Intrusion Detection Method Based On Machine Learning

At present,computer network technology is gradually mature and widely used,and the subsequent network security issues have also attracted people's attention.Intrusion detection technology can solve network security problems to a certain extent.Traditional intrusion detection technology relies on the completeness of the knowledge base or the pros and cons of the compilation of rules.The current commonly used machine learning methods can automatically learn useful knowledge from flow data more effectively.In particular,the popular deep learning has a more advanced ability to automatically mine "structures".Therefore,network intrusion detection technology still needs to do a lot of research and analysis in terms of intelligence and efficiency.This thesis analyzes the existing methods of network intrusion detection,aiming at the problems encountered in the intrusion detection process,design and implement the network intrusion detection model from the following aspects.First,preprocess the KDD99 data set,including operations such as normalization,one-hot encoding,and null value processing.In order to solve the situation that network flow data usually has category imbalance,this thesis implements a combination of Smote oversampling and Ensemble integrated undersampling to solve the problem of data imbalance.Second,the deep learning model based on Deep Belief Networks(DBN)and Stack Sparse Autoencoder(s SAE)and the Random Forest(RF)model are implemented based on the idea of integration,then the real data KDD99 is detected by three weak classification models.Finally,the majority voting algorithm proposed in this thesis and the weighted strategy of F1-score matrix are used to vote the prediction results of the weak classifier to get the final data category.Finally,the integration model proposed in this thesis is compared with the current popular algorithms such as decision tree and KNN,and the PCA dimension reduction method and the deep learning method in this thesis are used to solve the problem of feature redundancy.The experimental results show that the method in this thesis performs well in the detection of network attack types.Finally,this thesis applies the model to the intrusion detection system.The system can identify simulated DOS attacks,which reflects the practical value of this model.