CAPTCHA Image Generation using Style Transfer Learning in Deep Neural Network

Abstract

CAPTCHA is widely used as a security solution to prevent automated attack tools on websites. However, CAPTCHA is difficult to recognize human perception when it gives a lot of distortion to have resistance against the automated attack. In this paper, we propose a method to deceive the machine while maintaining the human perception rate by applying the style transfer method. This method creates a style-plugged-CAPTCHA image by combining the styles of different images while maintaining the content of the original CAPTCHA sample. We used 6 datasets in the actual site and used Tensorflow as the machine learning library. Experimental results show that the proposed method reduces the recognition rate of the DeCAPTCHA system to 3.5% while maintaining human perception.