Breaking 2D CAPTCHA Via Deep Learning And Designing Novel 3D CAPTCHAs

CAPTCHA is a security verification mechanism that distinguishes people from machines.It can effectively prevent malicious attacks of computer programs.Text CAPTCHAs and image CAPTCHAs are the most widely used verification mechanism at present.Currently,the research and analysis of these two types of CAPTCHA are relatively mature,and the security verification of CAPTCHA has been greatly tested.Therefore,a new type of verification mechanism with better security performance needs to be designed.This article mainly analyzed and researched the CAPTCHA from attack aspect and defense aspect,and obtains the following four results.Firstly,a recognition method of image CAPTCHA based on deep learning was proposed.Deep learning technology has a very significant advantage in image recognition.Convolutional neural networks can automatically learn and extract image features by training a large amount of image data.We analyzed the characteristics of the image CAPTCHA of the 12306 official network and arranged the 12306 image training data set.And then,we used the classic model of the depth learning in the image recognition which called Alex Net to classify the 12306 images.Our experimental results showed that the image CAPTCHA is not secure enough and may cause difficulties for human identification and verification.So we analyzed the structure of the 12306 image CAPTCHA,and then wrote a Python script to break down it with a recognition accuracy rate of 91%,which is much higher than human recognition accuracy.Secondly,we proposed a design scheme of gyroscope 3D CAPTCHA.Our verification mechanism is different from the traditional 2D CAPTCHA by using the mobile gyroscope additionally.Our experiment generated a CAPTCHA model through 3D modeling first,and then designed a random algorithm to switch CAPTCHA in real time,and implemented a gyroscope 3D CAPTCHA at the mobile terminal finally.The gyroscope 3D CAPTCHA is a verification mechanism based on visual perception.The user needs to adjust the angle of the mobile phone according to the visual perception and use the rotation of the gyroscope to find a suitable angle for identifying the CAPTCHA.At the end of our experiment,the efficiency and availability of the CAPTCHA are analyzed by manual tests.The results of the tests showed that the CAPTCHA has high practical value in actual application scene.Finally,we proposed a scheme of augmented reality CAPTCHA.By combining AR technology and 3D CAPTCHA,we improved the security strength of CAPTCHA.In this thesis,we studied the 3D registration technology and recognition tracking technology which are the key technologies in AR.The main function of 3D registration technology is to detect the orientation of the camera in real time and display the camera image to the correct position of the image.The function of the recognition and tracking technique is to track and recognize the target image and achieve the real-time fusion of virtual and real objects.Our experiment designed a random handoff algorithm for AR CAPTCHA,and implemented AR CAPTCHA on the mobile phone.User authentication needs to scan the PC image with the mobile phone camera,which can be observed at 360 degrees.Finally,the stability and security of the CAPTCHA were analyzed through psychological experiments and natural scene text recognition.This thesis analyzed the CAPTCHA from attack aspect and defense aspect.In attack aspect: we used deep learning technology to attack the image CAPTCHA,and found the existence of security risks;In defense aspect: we designed and implemented the gyroscope 3D CAPTCHA and AR CAPTCHA.Our experimental results showed that these two verification mechanisms have high availability and security,and have wide application prospects.