Research On Intrusion Detection Of Small Samples Based On Deep Learning

With the rapid development of science and technology,the Internet has made it convenient for people,but it has also brought a great threat to the security of cyberspace For example,the emergence of network viruses,hackers and other intrusions may lead to the loss of people or government property,and even cause serious social panic.Intrusion detection system as a security line in the computer can dynamically monitor abnormal traffic in the network.Therefore,it is of great economic and social value to study intrusion detection.At present,deep learning has become a common intrusion detection technology,and has achieved good research results in accuracy.However,there are still some shortcomings: the overall detection rate of the existing deep learning intrusion detection model is high,but the detection rate of the small intrusion samples in the data set is low,which is prone to over-fitting and has high false positives and missing alarm.In addition,the deep learning intrusion detection model has many parameters and long training time.Based on the above deficiencies,this paper conducts research at the data level and algorithm level respectively.The main research contents and contributions are as follows:(1)Aiming at the problems of low detection rate of small intrusion samples and slow training speed of deep neural network.This paper proposes an intrusion detection algorithm based on improved oversampling and deep learning model at the data level.First proposed to improve the sampling KSMOTE algorithm,because in the process of the traditional sampling without considering the sample distribution deal with the problem on the fringes of outliers,according to the calculation near neighbor samples of each sample balance rate and set up a threshold,depending on the position of sample distribution,the sample is divided into dangerous border data,data and security data types.Different oversampling operations are performed on different types of samples.In addition,considering that the existing deep learning intrusion detection models have many parameters and long training time,this paper proposes to decompose the traditional convolutional kernel into asymmetric convolutional kernel structure to reduce the parameters of the model,and proposes an asymmetric convolutional self-encoder(NCAE)to extract features from the data.Finally,an integrated twin-support vector machine classifier based on partial binary tree is constructed to detect the features.By combining the deep and shallow learning techniques,the accuracy of intrusion detection is guaranteed,the training time of the model is shortened,and the accuracy of small invasion samples is improved.(2)In the data set that is not conducive to sampling,there are some problems such as low detection rate of small intrusion samples,high rate of missing alarm and false alarm,and easy over-fitting in the process of model training.In this paper,a CNN(Convolutional Neural Network)and LSTM(Long Short-Term Memory)based deep pseudo siamese network intrusion detection algorithm is proposed.It is proposed to construct two different branch network structures on the structure of siamese neural networks: one is CNN branch network,which can effectively learn feature information in data;the other is LSTM branch network,which can combine features of different dimensions into sequences for input,so as to improve the generalization ability of the model.Through experimental verification,this model can improve the detection rate of small intrusion samples in the data set,improve the overall accuracy of the algorithm,achieve the research objectives of this paper well,and further improve the network intrusion detection technology.