Deep Learning-Based Text-based CAPTCHA Anti-Recognition Technology Research And Prototype Implementation

Text-based CAPTCHAs appeared early,has strong scalability,and is widely used.However,with the development of deep learning technology,the security of text-based CAPTCHAs has been increasingly challenged.Therefore,in order to enhance the antirecognition ability of text-based CAPTCHAs when facing deep learning models,this thesis proposes two types of deep learning-based text-based CAPTCHA anti-recognition algorithms suitable for different scenarios.These two types of algorithms can maximize the anti-recognition ability of original text-based CAPTCHAs without affecting the success rate of human eyes recognition.Firstly,study how to improve the anti-recognition ability of text-based CAPTCHAs as much as possible when computing resources are limited.Deep neural networks are very sensitive to adversarial samples.The adversarial samples generated by adding small disturbances that are difficult for humans to perceive can effectively fool the neural network models and make the models output completely different results.This is exactly in line with the original intention of the CAPTCHAs design.Based on this,this thesis proposes four adversarial text-based CAPTCHA perturbation algorithms in different application scenarios to enhance the anti-recognition ability of the original text-based CAPTCHAs.This algorithm does not need to collect a large number of samples for model training.A small range of targeted pixel changes can effectively interfere with the deep neural models,which is suitable for resource-constrained situations.Secondly,study how to maximize the anti-recognition ability of text-based CAPTCHAs when computing resources are sufficient.This thesis proposes a style transfer transformation network that can generate text-based CAPTCHA images with any style in real time.Just input an original text-based CAPTCHA image to be enhanced and a randomly selected style image into the trained network,an arbitrary style enhanced textbased CAPTCHA will be generated soon.The style transfer technology can well retain the character outline of the text-based CAPTCHAs.By continuously changing the style image,text-based CAPTCHAs with different styles can be generated,which greatly increases the diversity and randomness of text-based CAPTCHAs,without affecting the success of human eyes recognition.Therefore,the anti-recognition ability of original textbased CAPTCHAs is greatly enhanced.As it involves model training and requires a large number of training samples,this method is suitable when the training samples and computing power are sufficient.Thirdly,analyze the performance of the two types of text-based CAPTCHA antirecognition algorithms proposed in this thesis.This thesis uses 4 common CNN models as the anti-recognition ability evaluation model.By calculating the recognition success rate evaluation index,it detects the anti-recognition ability of the CAPTCHA images processed by the above algorithms with and without image preprocessing,including the change of the evaluation index of the CAPTCHA images before and after the algorithm processing by the anti-recognition ability evaluation model,and the interference effect of the processed CAPTCHA images on different recognition models.In addition,considering that the attackers may re-collect the text-based CAPTCHA samples,manually mark the samples and then train the model to improve the accuracy,this thesis simulates the situation and conducts a research on the suppression effect of the enhanced text-based CAPTCHAs generated by the anti-recognition algorithms proposed in this thesis on the retraining attacks.Fourthly,study the usability of enhanced text-based CAPTCHAs generated by the two types of text-based CAPTCHA anti-recognition algorithms proposed in this thesis.This thesis uses the Amazon Mechanical Turk crowdsourcing platform to test the usability of the original text-based CAPTCHAs and enhanced text-based CAPTCHAs,observe the changes in the success rate of human eyes recognition and the average recognition time,and test how much the algorithms in this thesis will affect the usability of the original text-based CAPTCHAs.Fifthly,in order to test the anti-recognition performance of the text-based CAPTCHA anti-recognition algorithms proposed in this thesis more intuitively and conveniently,this thesis designs and implements a text-based CAPTCHA anti-recognition platform,which integrates the two types of anti-recognition algorithms proposed in this thesis.Users can upload text-based CAPTCHA image to be enhanced on the platform to generate the corresponding enhanced text-based CAPTCHA image.At the same time,Users can observe the changes in the recognition results,confidence and other indicators after the text-based CAPTCHA is enhanced.The experimental results show that the two types of deep learning-based text-based CAPTCHA anti-recognition algorithms proposed in this thesis can effectively enhance the original text-based CAPTCHAs against deep neural network recognition without affecting human eyes recognition.