Intrusion Detection Based On Imbalanced Classification

With the development and popularity of network,the Internet has become one of the important components of society,however,with it comes the increasing problem of network security.In the complex network environment,network attacks may threaten the public security of society and cause serious consequences,and how to ensure network security has become an imminent problem.Intrusion detection system can actively detect complex attacks and threats,and it is an important line of defense to ensure network security.With the continuous updating of network attacks,the drawbacks of traditional intrusion detection systems come out gradually,such as low detection efficiency and poor performance.In addition,traditional intrusion detection systems assume that data is balanced and focus only on overall accuracy,resulting in a lack of attention to minority samples.However,the real network traffic distribution is extremely imbalanced,i.e.,the proportion of abnormal data in the network data is very low,which leads to the underreporting of attacks.To address these problems,an intrusion detection model based on imbalanced learning and feature extraction is proposed in this paper.The main work is as follows:For the imbalanced classification problem,a Hybrid Sampling Method Based on Stratified Sampling and Moth-flame Optimization is proposed to deal with the imbalanced data.First,the skewed decision boundary of the original imbalance data is obtained using support vector machine(SVM).Then,the majority samples are stratified undersampled,focusing on removing the majority samples near the boundary to make the boundary clearer.The minority samples are oversampled and new samples are generated near the boundary based on the support vector and surrounding samples.Since the boundary vicinity is sensitive,to avoid the generation of noise,the Moth-Flame Optimization Algorithm(MFO)is used to optimize the new samples to finally obtain balanced data.Experiment is conducted on 29 imbalanced datasets provided by KEEL data repository,and G-mean and AUC are used as evaluation metrics.The experimental results show that the hybrid sampling method can effectively improve the classification performance of the imbalanced data.To address the problem that traditional intrusion detection systems are difficult to extract high-dimensional data,which leads to low detection rate,a feature learning method based on CNN and GRU is proposed in this paper,using Convolutional Neural Networks(CNN)to extract spatial features and reduce feature dimensionality,and using the Gated Recurrent Unit(GRU)to extract information features with large temporal distance in the time series to achieve comprehensive and effective feature learning.A hybrid sampling method based on stratified sampling and moth-flame optimization and CNN-GRU model are combined to propose an intrusion detection model based on imbalanced learning and feature extraction.Experiment is conducted on the UNSW-NB15 dataset and four performance metrics,accuracy,precision,recall and F1-Score,are used as model evaluation metrics.The experimental results show that the proposed intrusion detection model combined with imbalanced learning and feature extraction can effectively improve the intrusion detection performance.