Web attack detection technology is an important technology to protect computer web applications.With the continuous development of web attack technology in recent years,web attack detection technology faces some major challenges.For example,the traditional blacklist technology security protection measures are not flexible enough.And the machine learning technology based on Feature Engineering is time-consuming and laborious,the detection rate and accuracy are low,and the false positive rate is high.Deep learning,because of its powerful feature extraction and feature expression capabilities,greatly enhances the performance and application range of deep learning,making deep learning a new way to detect Web attacks.The main work of this paper is as follows:(1)This paper proposes a method to construct the word vector of HTTP request data based on word embedding.In view of the problems of high dimension of one hot coding feature vector,inability to maintain the original semantic features of text,and low quality of Web attack features,a method of using word embedding is proposed.This method first extracts the initial features of HTTP request data,and then uses word2 vec based on word embedding to transform HTTP request data into word vector,the obtained word vector greatly reduces the dimension using one-hot encoding and maintains the original semantic features of the HTTP request data.Finally,the obtained word vector is projected into the two-dimensional space.Experiments show that the method can extract the features of Web attack efficiently and accurately.(2)A Web attack detection method based on Bi-directional Long Short-Term Memory is proposed.Aiming at the problems of time-consuming and low performance of Web attack detection based on traditional machine learning,as well as the problems of less research on Web attack detection based on deep learning,low detection rate and accuracy rate,and not giving full play to the effect of end-to-end detection of deep learning.This paper studies the current research status of network intrusion detection using deep learning,and proposes a method of using Bidirectional Long Short-Term Memory Network for Web attack detection combined with the characteristics of HTTP request data in Web attack detection.The experiment proves that the detection model proposed in this paper can detect Web attacks well,and the detection rate and accuracy are improved to a large extent while maintaining a low false positive rate,and achieve the end-to-end detection effect.(3)This paper proposes a new method based on probability model to generate text Adversarial Examples.Firstly,this paper introduces the current phenomenon of Adversarial Examples in the field of computer vision,and investigates the cases of Adversarial Examples in the field of malware detection.Combining the research status of Adversarial Examples in the field of text classification,this paper analyzes the security problems of the current Web attack detection model based on deep learning,and puts forward the generation method of Adversarial Examples based on probability model,which is proved by experiments.The method proposed in this paper can effectively improve the generation efficiency of text countermeasure samples,and provide a certain reference value for the follow-up better defense of this kind of security problems,and improve the robustness of Web attack detection model based on deep learning. |