Design And Security Analysis Of New CAPTCHAs Based On Image Style Transfer

CAPTCHA is a Turing test program designed to fully distinguish between humans and machines.Over the last years,CAPTCHA has been proved to be an effective way to resist malicious attacks on automatic script programs.Therefore,CAPTCHA is being widely used in Internet websites to prevent malicious registration,malicious comments,malicious postings,and violent password recovery.In the scenario,a low cost method is used to maintain network security and play an increasingly important role in network security.A well-designed CAPTCHA means that it has good security and ease of use.However,it is difficult for CAPTCHA designer to reach a balance between good usability and high security.Although the complex CAPTCHA improves the security of the CAPTCHA,the user is more complicated to use,and it takes more time to pass the challenge.With the development of deep learning technology,the security of CAPTCHA also faces challenges.In this paper,we will use Neural Style Transfer,a new deep learning technology,to enhance the security design of the CAPTCHA without affecting the usability.The main work of this paper is as follows:(1)An end to end real-time Neural Style Transfer network has been designed and improved to achieve a 112*112 pixel size image style transfer within tens of milliseconds.Compared with the early Neural Style Transfer network,it has thousands of times of speed improvement.One of the important factors that make CAPTCHA easy to crack is that there are fewer data sets,and an attacker can easily access almost all samples.However,after the real-time image style transfer,CAPTCHA can be generated in real time once each verification request is received.Different styles can be used to generate different styles of CAPTCHA for the same image,and the CAPTCHA data set is increased exponentially.(2)Three new image CAPTCHA based on neural style transfer are designed,and detailed user experiments are carried out.The three CAPTCHAs include: Grid-CAPTCHA SlidingCAPTCHA and Font-CAPTCHA.The Grid-CAPTCHA provides 9 stylized images,and users need to select all the corresponding images according to a short text description.The average time of verification is 11.83 seconds,and the accuracy rate is 75.04%.The SlidingCAPTCHA requires the user to slide the jigsaw fragments to the specified shadow position.The average time of verification is 1.48 seconds,and the accuracy is 88.12%.The Font-CAPTCHA prompts the user to click on the Chinese characters appearing in the prompt in the order of the text description.The average verification time of the user is 4.44 seconds,and the accuracy rate is 84.49%.By verifying the time and accuracy,we can see that the two new CAPTCHAs designed in this paper have high usability.(3)In order to verify the influence of traditional image processing and deep learning technology on the security of the CAPTCHAs,this paper makes a comparative analysis of the security of the new CAPTCHAs and other similar verification mechanisms.The contrast experiment results show that the Neural Style Transfer technology applied in the CAPTCHAs can effectively reduce the success rate of automatic attack.At the same time,the experimental results also show that the deep learning technology can not only be used to crack the CAPTCHAs,but also has a positive effect on the improvement of the security of the CAPTCHAs,and provides a promising direction for the future research of the CAPTCHA.