Secured Text Based CAPTCHA using Customized CNN with Style Transfer and GAN based Approach

A CAPTCHA is one of the popular, secure methods for securing web content from unauthorized access or malicious intruders. It ensures that the user is not a malicious user but a human being. In the early phase of the CAPTCHA scheme, text-based CAPTCHA was widely used, but with recent advancements in deep learning, it has become vulnerable. For this reason, many organizations use image-based CAPTCHA or puzzles to secure their systems. However, these image or puzzle systems sometimes take a lot of time to give access to the system when the user may not understand clearly the images or puzzles. As a result, users sometimes feel bored accessing the systems as it takes a long time. So, considering this scenario, a text-based CAPTCHA has promising demand in terms of usability concerns. The deep learning model can be applied to make complex backgrounds that are hard for machines to understand but easy for human beings; thus, text-based CAPTCHA will be a more secure method compared to images or puzzles. In this paper, we use the style transfer CNN and GAN based deep learning model to make the complex background of the text-based CAPTCHA. Our proposed model can secure a text-based CAPTCHA, and the experimental results show good performance. We try to show the vulnerability of text-based CAPTCHA when we don't use style transfer GAN, and it shows the CAPTCHA recognition rate is near 98.68%. But, with our proposed model, the CAPTCHA recognition rate is 2.1 %, which is a promising result for using text-based CAPTCHA in security systems.