| With the rapid development of the Internet,a variety of Web applications and Web services are emerging,which make people’s daily life more convenient.However,the cyberattacks targeting at Web servers are also increasing.Web applications and Web services are facing cyber threats such as malicious request attacks,which cause damage to the property and privacy of the companies and users.Therefore,how to detect the malicious request attacks quickly and accurately,and prevent them from executing are quite important.In the field of malicious request detection,the rule-based methods require a lot of expert knowledge and manpower to design and maintain the rule library;the statistical analysis based methods and machine learning based methods also need to manually design and extract features from the raw requests,which are quite tedious.To solve these problems,this thesis considers a HTTP request message as a string,and transforms the task of malicious request detection as text classification.This thesis proposes a character-level embedding with convolutional neural network(CNN)model for malicious request detection.Specifically,this model first utilizes Word2 vec to capture the semantic information in HTTP request and learn its character-level embeddings.Furthermore,a specially designed CNN is applied to automatically extract high-level features from the character-level embeddings for malicious and legitimate request identification.In particular,this thesis proposes a multiscale feature fusion approach for malicious request detection to improve classification performance by exploiting the rich semantic information contained in HTTP request.Firstly,it models the HTTP request in both characterlevel and word-level simultaneously.Secondly,it extracts the high-level semantic features in HTTP request with a CNN based network.Thirdly,it jointly learns the multiscale representation for HTTP request with the help of multimodal learning techniques.Finally,a linear classifier is adopted for classification.In order to show the effectiveness of these two proposed methods,a series of experiments have been conducted on a widely-used dataset and an authentic network traffic dataset.Experiment results show that the proposed approaches have significant improvement on the performance against existing state-of-the-art methods,and are able to learn discriminative representations for HTTP requests. |
