Research On Deep Learning Adversarial Attack Oriented To Intrusion Detection

With the continuous development of the society,the Internet has gradually penetrated into every aspect of life,and people pay more and more attention to network security.In recent years,network security-related events keep appearing,which have a huge impact on our lives,making people realize that network security has become an important factor of social stability,and any network security problem may cause disastrous consequences.Intrusion detection is an important means to ensure network security,is an effective defense technology,has an important role in network security.With the advent of the era of big data,traditional intrusion detection technology based on rule matching,statistics and other methods can no longer perform well in the face of massive,complex and unbalanced network data.Now,improving the overall performance of intrusion detection is an important task in the field of network security.The development of deep learning provides a new opportunity for the development of intrusion detection,but at the same time,the potential security problems of deep learning should be prevented.In order to improve the efficiency of intrusion detection,deep learning technology is used to detect network data.However,deep learning itself has security problems and is vulnerable to counter attacks.This paper proposes a black-box attack method based on the transferability of deep learning,providing a new idea for the development of intrusion detection and network security.The main research work of this paper is as follows:First of all,this paper studies the working principles of the multilayer perceptron,deep convolutional neural network and long and short memory network,and then puts forward the general framework of deep learning intrusion detection.According to the input characteristics of each deep learning model,the data set of NSL-KDD is individually coded and normalized to meet the input requirements.Deep learning model and traditional machine learning model are built to train and extract features,classify network data,and realize the purpose of intrusion detection.Through experiments,it is proved that deep learning is obviously better than the traditional method when dealing with intrusion data.Secondly,this paper studies the feasibility and significance of using adversarial examples to attack the intrusion detection model on the basis of three deep learning intrusion detection methods.According to whether the attacker has the information of the target model and data set,the attack can be divided into white box attack and black box attack.In general,JSMA algorithm with target attack and FGSM algorithm without target attack are adopted in the implementation of white-box attack.Using the transferability of the adversarial examples,that is,the adversarial examples cross-model can still complete the attack,so as to realize the black-box attack of the intrusion detection model.This attack can be achieved against the attack without reducing the background knowledge of the opponent and knowing the structure of the target model.Through the experiment,the black-box attack on the deep learning intrusion detection model by using the adversarial examples is proved to be effective.