Abstract:
This project borrows techniques from computer vision to build new information security tools and primitives. The issues of website attacks like DOS attacks have been common in the past decades. To cater this issue CAPTCHAs were developed. Initially text-based CAPTCHAs were in use but since the development of machine learning algorithms and advanced OCR, text-based CAPTCHAs started to fail, and their use decreased.
The aim of the project is to develop new text-based CAPTCHAs by applying various techniques on text and engraving the text in natural scene images, making it undetectable by state-of-the-art detectors while it remains easy to read and solve by the humans.
The goal is to make the text using different objects and place it in different scenes. There are infinite objects, so even if some detectors manage to learn the few objects that are used and can detect the text, the objects can be changed and the detectors will have to learn again to detect the text made from the new objects, keeping the problem complex for the detectors.
Currently Google’s re-CAPTCHA is in use. It has a good security but some usability issues because of which it is not the ideal CAPTCHA. Humans find easy to read text rather than selecting boxes comprising of certain thing. That’s why text-based CAPTCHAs are preferred over others.
