Abstract: CAPTCHA is widely used in the registration and login process of websites and applications to distinguish normal users from programs.However,with the advancement of deep learning,many deep learning recognition methods for CAPTCHA have been proposed.CAPTCHA can no longer distinguish human users from computer programs effectively,and its security has been greatly challenged.The adversarial example can make the output result of neural network completely different from its original predicted result.Recent researches find that combining adversarial example with CAPTCHA is an effective method to resist the attack of deep learning recognition system.Researchers use adversarial example generation methods to generate adversarial chara-cter CAPTCHA to defend against deep learning methods.Existing adversarial character CAPTCHA generation methods are white-box methods that require knowledge of the structural parameter information of the attacking network.However,practical CAPTCHA application scenarios usually do not know the information of the attacking network,so robust CAPTCHA should be able to perform well without knowing the attack information.In this paper,a character-based adversarial CAPTCHA generation method(ACoDE) based on differential evolution algorithm is proposed to improve the solving ability of the algorithm by optimizing the scaling factor in the mutation process and the population evolution strategy.Without knowing the information of the attacking network,the adversarial examples generated by the proposed method are more capable of misleading the neural network.The adversarial example generation method is used for the character CAPTCHA dataset,and the success rate of the current state-of-the-art CNN character-based CAPTCHA recognition system reduce to less than 30%.The visual effect of the adversarial CAPTCHA is satisfactory when compare with other white-box methods.

Key words: Deep learning, Adversarial examples, Differential evolution algorithm, CAPTCHA, Network security