Automated CAPTCHA Generation from Annotated Images Using Encoder Decoder Architecture

Abstract

A CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) is a program that protects websites from bots by generating and grading assessments that humans can pass but current computer programs cannot. CAPTCHAs provide security from bots in applications such as preventing comment spam in blogs, protecting website registrations etc. Recent breakthroughs in Artificial Intelligence (AI) have led to development of systems which can crack current image based CAPTCHAs with over 90% accuracy. This necessitates the need for new types of image based CAPTCHA which would be plausible for a human to solve it, and, at the same time, pose a challenge for a bot to break the CAPTCHA. To this end, the paper proposes a novel type of image based CAPTCHA where the user is presented with a set of images and a multiple answer based question based on the contents of the image-set. The question generated is such that a human is able to answer the question easily, whereas a bot would have to delve into the intricacies of image recognition, natural language processing on the question and then perform a knowledge correlation with the options to crack the CAPTCHA which is a rather tedious task to achieve. The novelty in the CAPTCHA presented can be expressed in terms of the CAPTCHA type itself as well as the deep learning architecture employed to synthesize the CAPTCHA. The proposed CAPTCHA generation system uses an Encoder Decoder architecture whose basic building block is a Gated Recurrent Unit (GRU) - a type of Recurrent Neural Network (RNN). The proposed system also facilitates a dynamic CAPTCHA generation mechanism eliminating the need to store a mapping between the images and questions.