| As the first line of defense of Internet security,CAPTCHA provides a guarantee for mankind to continuously benefit from the convenience of the Internet.But the continuous development of deep learning in recent years has unexpectedly created conditions for hackers to crack CAPTCHA maliciously,bringing great threats and challenges to the security of CAPTCHA.In order to solve the problem described above,CAPTCHA designers are constantly raising new CAPTCHA mechanisms,but the emerging CAPTCHA’s complex operation logic reduces its user-friendliness.Fortunately,recent researches on adversarial attacks have provided new ideas for CAPTCHA designers.It has been found that the neural network models are still fragile.Models can be misclassified by adding small noise that is not detectable by humans easily.This property of adversarial examples just meets CAPTCHA’s requirements of the security and usability,and the adversarial CAPTCHA emerges as the time require.Based on the above considerations,this paper proposes two adversarial CAPTCHA generation algorithms,whose essence is adversarial attack algorithms.The main research contents of this paper are as follows:(1)We propose an adversarial CAPTCHA generation algorithm based on flipping and rotation strategies.Utilizing the flipping and rotation operations,this method overcomes the shortcoming of iterative fast gradient sign method that is easy to overfit,and then improves the transferability of adversarial examples in black-box setting.Adam gradient descent method combined with auxiliary loss is used to find a more accurate gradient direction and improves the attack success rate without adding excessive perturbations.(2)We propose an adversarial CAPTCHA generation algorithm based on variable step size and filtering strategies.This method solves the problem that the algorithm may fall into a local optimum when using the transferability of the surrogate model for black-box attack.The filtering operation maximizes the retention of the perturbations that contribute the most to attack effect,and prevents image preprocessing techniques from eliminating adversarial perturbations,which improves the transferability of adversarial examples under the ensemble adversarial training models and reduces the visibility of noise.In addition,the algorithm also uses the adam method combined with auxiliary loss to make gradient direction clear,and shorten the generation time,having practical significance.(3)We propose a new evaluation metric of adversarial attack,that is the contribution for reducing perturbation size.This metric takes both the attack success rate and visibility of noise into consideration,which is more convictive.(4)We use the Vue and Spring Boot frameworks to build a platform to demonstrate the process of generating adversarial CAPTCHA and test its performance.The proposed algorithms are validated on Image Net-compatible dataset in terms of transferability,visibility,contribution for reducing perturbation size and generation efficiency.Experiment results show that the two adversarial CAPTCHA algorithms presented in this paper perform well,and can generate adversarial examples with higher transferability and lower noise visibility,which meet the requirements of security and usability of the CAPTCHA.Finally,the security and usability of adversarial CAPTCHA are tested and analyzed through experiments and user surveys on Image Net dataset,which provides a reference for the actual online of CAPTCHA. |
