Research On Malicious URL Detection Based On Generative Adversarial Networks

Malicious web pages,which seriously affect people's property security and personal privacy issues of network security,it is urgent to solve.Generally speaking,the life cycle of malicious web pages is short,and it is difficult to collect and label the data will make the existing methods unable to detect the latest malicious domain name in real time.In addition,the detection efficiency of traditional detection methods such as machine learning is greatly affected by the accuracy of data sets and the construction of artificial features.In respect of the issues above,the main innovations in this article include the following three points:1.Designed URL character-level encoder.We want to use the way that deep neural networks can automatically extract features to improve the machine learning sensitive to feature engineering in identifying malicious URLs,After analyzing the characteristics of the malicious URLs,this paper designs a character-level encoder,which uses the encoded data as input,trains the Generative Adversarial Networks to Generate data that approximates the real sample.2.Generate malicious urls to expand the dataset.Because of the model can fit the distribution characteristics of real samples.It is used as an extended data set to train the traditional classifier,and the accuracy of the classifier is improved after adding the generated data through experimental comparative analysis,so as to prove the validity of the generated data and solve the problem of insufficient samples.3.Propose the Mu-GAN model to detection malicious web page.On this basis,the original adversarial model was improved by adding another discriminator is trained to be a classifier that can distinguish between malignant and benign URLs.Finally,a comparative experiment shows that the accuracy of the model in this paper is the same as of state-of-the-art supervised machine learning model.These results show that this model can complete the task of classifying malicious URLs.In fact,our experiments show that the use of two discriminators can provide high stability,which is a well-known problem with the GAN architecture.Based on GAN,this model is an innovative application of malicious URL classification,which provides a reference for the application of GAN in network security,and also gives a new way to solve the text problem.