Research On Network Intrusion Detection Method For Data Dimensionality Reduction And Data Imbalance

With the continuous development and popularization of the Internet,the corresponding network attacks are also increasing.Cyber attack has become a behavior that threatens people’s property,privacy,personal information and national security.It will not only cause losses to enterprises,government agencies and other organizations,but also lead to the impact and imbalance of social public interests.However,in today’s network environment,various attacks are constantly updated,and more and more network attacks threaten data security.Traditional intrusion detection systems are greatly challenged.Intrusion detection is an important technology to maintain network security,and has been widely studied by scholars.In the field of intrusion detection,data usually has high dimensionality and complexity.Data dimensionality reduction can reduce the amount of calculations,remove redundant information,improve visualization and model generalization capabilities.However,the data imbalance leads to a challenging problem that there are far more normal data than attack data.The imbalance of the data can lead to the bias of the decision boundary,which leads to the misclassification of higher value attack data.In the face of imbalanced data,how to make the classification model to classify more effectively is called the imbalanced learning problem.The work done in this research is as follows:(1)Aiming at the problems of high dimensionality of network intrusion detection data,data redundancy,high false alarm rate,and low detection rate,this paper first proposes a method for parallel feature extraction of CNN and GRU,using multi-head attention mechanism to combine CNN and Bi GRU The extracted spatial and temporal features are fused,trained with a gated recurrent unit model,and finally classified.Further,in order to solve the data dimensionality reduction,the improved stacked sparse denoising autoencoder is used to optimize the model.By including the information item reflecting the attribute feature relationship as a penalty item in the loss function,the local approximation and decoding performance of the autoencoder network can be improved.performance in terms of ability.(2)Aiming at the problem of data imbalance,this paper proposes a mixed sampling algorithm called KMSMOTE.The algorithm uses different sampling strategies to deal with majority class samples and minority class samples.For the minority samples,the SMOTE oversampling algorithm is improved,and the decision boundary samples and safety samples are sampled according to different thresholds,and the noise samples are not processed,which solves a series of problems caused by oversampling.When dealing with majority class samples,K-means clustering and proportional random sampling are adopted to achieve the purpose of data balance.The experimental results show that the improved autoencoder proposed in this paper has better performance than the ordinary autoencoder,and the parallel data processing method has better performance than the serial method.The boundary overlap and noise problems caused by the SMOTE algorithm are improved by KMSMOTE mixed sampling.