Application Of Deep Learning In Web Security Detection Scenario

In recent years,with the deepening of information technology and the rapid development of big data,cloud computing and other technologies,more and more government units and private enterprises choose to move toward "Internet +".A large number of e-government,inclusive finance,online games,live broadcasting with goods and other Web services have also been online.But at the same time,while providing convenient services,enterprises' Web applications have been subjected to a large number of network attacks from domestic and foreign attackers.Attacks against Web services such as SQL injection,XSS attack,Webshell implantation,directory traversal and so on are common.However,in the face of increasingly complex attack threats,the traditional Web security detection methods mainly use statistics for identification.Such methods have problems such as low recognition rate and difficulty in detecting unknown threats.To solve this problem,deep learning technology is introduced in this paper to improve the detection accuracy and adapt to the Web detection of unknown threat types.The main work of this paper is as follows:First of all,this paper analyzes the common Web attack methods,and mainly standardizes the URL data of Web request,then uses syntax and semantic analysis methods to make vector representation of it.Finally,after comprehensive evaluation,Word2 VEC model was used to represent word vectors,and parameter matrix was obtained by unsupervised training.This model can retain and extract the effective feature information in the URL to the greatest extent,and is a vector representation model more suitable for the goal of this experiment.Secondly,this paper focuses on the classification detection algorithm based on deep learning,and employs LSTM(Long Short-Term Memory)and Text CNN(Text Convolutional Neural Network)algorithms to train and detect the samples.On this basis,according to the respective characteristics of the two algorithms,the combined model training method based on CNN and LSTM is proposed.Firstly,the local features of URL data are extracted by using CNN model,and then the LSTM model is used to combine the features of different dimensions into sequences for input,as to solve the timing dependence problem in the data.The experimental results show that the accuracy of this model is 99.1%,and the false alarm rate and missing alarm rate are obviously higher than the other two classification detection models.The experimental results show that this combination model can more effectively extract the hidden characteristics of Web attacks,and better achieve the research objectives of this paper.At the end of this paper,the content of the experiment is summarized and refined,the shortcomings of the experiment are analyzed,and the next step of the work and the future development trend of the technology are prospected.