Research On Technology Of Instrusion Detection Data Augmention Based On GAN

While the Internet has brought us convenience,network security issues have become increasingly prominent.Users' personal privacy and data have been repeatedly attacked,including botnets,web phishing,and hosting fraud.Traditional firewall defense technologies cannot quickly respond to complex and changeable attacks.With the rapid development of deep learning technology,intrusion detection technology with strong feature self-learning capabilities enables the defense system to actively and proactively warn against attacks.However,this data-driven deep learning intrusion detection technology faces problems such as lack of training samples and undisclosed private data,which seriously affects the development of intrusion detection technology.Further research is urgently needed at the level of data enhancement.Aiming at the problem that the intrusion detection system based on deep learning lacks diversity training data and it is difficult to identify evolutionary attacks,this paper analyzes the characteristics of intrusion detection data sets and studies the intrusion detection data augmentation model based on Generative Adversarial Networks to generate and predict attack behaviors.The research has achieved the following results:Aiming at the problem of the lack of evolutionary training data for intrusion detection systems based on deep learning,this paper proposes an intrusion detection data generation method based on conditional Wasserstein Generative Adversarial Networks.This method uses Auxililary Classifier Generative Adversarial Networks to generate intrusion detection data.The additional conditional loss judgment of the discriminator will correct the label information of the generated data.By introducing Wasserstein distance and gradient penalty terms,the problems of traditional model training instability and model collapse are solved.Experimental results show that the proposed method can generate high-quality attack samples according to the specified type.Aiming at the problem of poor generation of minority samples in the process of data augmentation,a non-functional feature perturbation generation model is designed for R2 L attacks.For this type of attack sample,first keep the functional features unchanged,then split the non-functional features and input them into the generation confrontation network,and finally concatenate the functional features and the generated non-functional features to output.Experimental results show that the proposed model can generate small batches of data.Aiming at the problem that traditional Generative Adversarial Networks are poor in processing discrete features and difficult to express features efficiently,this paper proposes a fusion cross-layer Generative Adversarial Networks to generate discrete and continuous features in parallel,using softmax and sigmoid activation functions to generate both.Taking into account the weak correlation between some features,in order to better model the relationship between the combined features,the cross-layer network is introduced,that is,the feature cross is applied at each network layer to automatically learn the combined features,which overcomes the inefficiency of manual feature engineering.The experimental results show that the proposed model can better deal with discrete features and generate high-quality samples.In order to verify the validity of the generated data and solve the problem that the existing research only uses a single classifier method for data evaluation,this article measures the quality of the generated data from multiple angles and multiple methods.The experimental results show that the generated data and the real data have similar feature distributions,and the model generalization ability of the enhanced data training is improved,which verifies the effectiveness of the model and that the generated data can be used to expand the real data set.