Research On CAPTCHA Technology Based On User Behavior Characteristics

With the popularization and development of Internet technology,the importance of information security has become increasingly prominent,and the threat of information security becomes more serious,which brings new challenges to network information security.As the portal to ensure the security of network environment,the research on verification code technology has the extremely importanttheoretical value andpractical significance.At the same time,with the development of Completely Automated Public Turing Test to Tell Computers and Humans Apart(CAPTCHA)crack technology,more and more researchersexplore new CAPTCHA to defend attack.Consider the research background of the network verification code,a new verification codebased on the user behavior characteristics withusing the mouse is presented.Different from the traditionalpicture CAPTCHA to verify the ability of user identification,behavioral CAPTCHA has dual authentication: the firstis the matching of the image,the second is based on the behavior feature extracted from the mouse track of the human computer interaction to tell humansand computers apart.Details are as follows:Firstly,the track of dragging mouse by user to complete the verification is analyzed and modeled.Based onthe mouse dynamics,the information Entropy,and time series of symbolic theory,the feature of user behavior is extracted.The extracted features include static characteristics such as the time and the displacement offset of acquisition point,and dynamic characteristics such as drag speed,etc.Then discretizing features aims to achieve stronger robustness.What's more,the relative entropy is used to measure the difference of features extracted from the two kinds of samples,so as to choose the effective features to tell the human beings and computerprogramsapart.Secondly,the extracted features can express the user behavior characteristics from different views.But its high dimensionleads to time-consuming.In this thesis,the input space is mapped into a low dimensional feature space through the Principal Component Analysis.In order to achieve the most distinguishing feature,the LinearDiscriminant Analysis is used to further extract the feature subspace,and the behavior identification model is established to tell the human beings and the computer programs apart.Thirdly,this thesis uses the Support Vector Machine(SVM)to tell the human beings and the computer programs apart.It's hard to distinguish the mixed sampleswithhuman beings and computer programs.To solve the problem,one class SVM is preliminary used for anomaly detection.The parameters ofhigh accuracy and the best performance of the model are obtained through the parameter optimization algorithm.This thesis researches the influence of parameter selection on the recognition rate of SVMbased on the extracted mouse behavior data.Experimental results show thatthe setting of kernel parameter gamma and the error penalty parameter Clargely influence the classification performance of SVM.Finally,in order to solve theproblem of classification deviation,from the view of the Riemannian geometry,based on meeting the Mercer theorem,conformal transformation is made on the Gauss kernelof SVM to modify the kernel function.The objective function is to maximum the classificationinterval,then through multiple iterative to train the classification model.Carried out on the data of the behaviorcharacteristics,theresults show that the modified kernel function improves the classification accuracy and reduces the number of support vector.The improved model has better generalization ability and the abandon rate of computers is improved with the ensured recognition rate of human beings.