Research On Intrusion Detection Algorithm Based On Adaptive Sampling And Improved Convolutional Neural Network

With the rapid iteration of new communication technologies,the Internet has been further integrated into people's lives.It widely serves military,medical,education,finance and other aspects.Internet technology has brought us a lot of convenience,but also a variety of network security problems,common hacker attacks,trojan virus,network crime,which seriously affect people's modern network life,so the network and information security issues become particularly important.Network intrusion is a common means of active attack,also known as network attack.Intruders destroy the network environment or steal users' private information by illegal means.Attack recognition mechanism is a self-defense strategy for early warning of possible attacks in modern network.It can effectively identify network anomalies and prompt users to take timely security measures to avoid network hazards.However,the diversity of attacks poses a severe challenge to the existing intrusion detection systems.Traditional attack recognition methods usually identify anomalies by mining the association between data,and usually have some disadvantages in real network environment,such as high false alarm rate(FAR),low recognition accuracy(ACC),poor generalization ability.To enhance the comprehensive performance of the intrusion detection system(IDS)and ensure network security,we propose a hybrid attack recognition model based on adaptive synthetic(ADASYN)sampling algorithm and upgraded convolutional neural network(CNN).The improved attack detection model can be applied to large-scale network environment with security risks,so as to achieve the purpose of taking precautions against abnormal network environment.Our main work is as follows:(1)Firstly,for the extreme imbalance of sample categories in the original data set,we use the ADASYN algorithm to augment the data sample.The balanced data set can effectively prevent the model from leaning large sample features while ignoring small samples in the process of model training,so as to improve the learning and expression ability of the model for small samples.(2)Secondly,aiming at the problem of single feature extraction scale and redundancy between channels of intrusion detection model based on traditional neural network,we design a new intrusion detection SPC-CNN model based on splitting convolution(SPConv)module of feature decomposition.On the one hand,the improved SPC-CNN model can increase the diversity of feature extraction in convolution process;on the other hand,different levels of convolution operations are used for feature map to effectively reduce the impact of information redundancy interchannels on model training.(3)Finally,An AS-CNN model mixed with the ADASYN algorithm and SPC-CNN is used for intrusion detection tasks,and the standard NSL-KDD data set is selected to test the effectiveness of AS-CNN model.The simulation results indicate that the recognition ACC of this model is 4.60% and 2.79% higher than that of the traditional CNN and RNN models,and the detection rate(DR)is 11.34% and 10.27% higher than that of the traditional CNN and RNN models,respectively.In addition,compared with the two models,the FAR is reduced by 15.58% and 14.57% respectively.The results indicate that the AS-CNN model proposed in this paper can effectively detect the common types of network anomalies and new types of network attacks in intrusion detection tasks,so it has good practical value.