Automation of complex text CAPTCHA recognition using conditional generative adversarial networks

Abstract

With the rapid development of Internet technologies, the problems of network security continue to worsen. So, one of the most common methods of maintaining security and preventing malicious attacks is CAPTCHA (fully automated public Turing test). CAPTCHA most often consists of some kind of security code, to bypass which it is necessary to perform a simple task, such as entering a word displayed in an image, solving a basic arithmetic equation, etc. However, the most widely used type of CAPTCHA is still the text type. In the recent years, the development of computer vision and, in particular, neural networks has contributed to a decrease in the resistance to hacking of text CAPTCHA. However, the security and resistance to recognition of complex CAPTCHA containing a lot of noise and distortion is still insufficiently studied. This study examines CAPTCHA, the distinctive feature of which is the use of a large number of different distortions, and each individual image uses its own different set of distortions, that is why even the human eye cannot always recognize what is depicted in the photo. The purpose of this work is to assess the security of sites using the CAPTCHA text type by testing their resistance to an automated solution. This testing will be used for the subsequent development of recommendations for improving the effectiveness of protection mechanisms. The result of the work is an implemented synthetic generator and discriminator of the CGAN architecture, as well as a decoder program, which is a trained convolutional neural network that solves this type of CAPTCHA. The recognition accuracy of the model constructed in the article was 63 % on an initially very limited data set, which shows the information security risks that sites using a similar type of CAPTCHA can carry.

Keywords

• text-based captchas
• deep learning
• conditional generative adversarial network
• cgan
• cnn
• information security