Research On Intrusion Detection Method Based On Inception-CBAM

With the popularization of the Internet,the security risks brought by network intrusion are constantly being exported,and the Internet has the problem of having a large amount of data and significant differences in the amount of data for intrusion attacks.Therefore,the accuracy of intrusion detection methods is the key to network security protection.In order to improve the accuracy of intrusion detection methods,an intrusion detection method based on Convolutional Neural Networks(CNN)Inception network structure and Convolutional Block Attention Module(CBAM)is proposed.First,in the data processing phase,the Borderline Smote oversampling algorithm is used to increase the number of sub class samples,and Min Max normalization is combined to process the data set,which can effectively alleviate the problem of large differences between the number of sample tags and the value of features;Secondly,an improved one-dimensional convolutional neural network Inception structure is used to extract features from different scales,and an improved method is constructed in conjunction with the CBAM attention mechanism to extract high robustness and low dimensional ensemble features;Finally,Adam optimization algorithm and Dropout regularization are used to improve the generalization of the model,so that the method has a better feature expression when dealing with massive data.We selected the classic intrusion detection dataset CIC-ISDS-2017 for simulation experiments,and selected four evaluation criteria: accuracy,accuracy,recall,and F1 value to comprehensively analyze the detection performance of intrusion detection methods.The experimental results show that the average accuracy of the Inception CBAM method reaches 99.57%,proving that the Inception CBAM intrusion detection method has good adaptability and high recognition ability for unknown samples.Compared with SVM,CNN,RNN,and BLS-GMM,the accuracy has been improved by 4.48%,1.35%,1.62%,and 0.04%,respectively,and the recall has been improved by 4.48%,1.36%,1.62%,and 0.14%,respectively.Meanwhile,compared to the comparative model,the detection rates of Bot,Web Attack Sql Injection,and Heartbleed for subcategory samples have been improved by 72%,65%,and 65%,respectively.The accuracy has been improved to some extent in cases of imbalanced data.The dissertation contains 14 figures,15 tables,and 57 references.