A Parametric Study of Arabic Text-Based CAPTCHA Difficulty for Humans

Abstract

The Completely Automated Public Turing test to tell Computers and Humans Apart (CAPTCHA) technique has been an interesting topic for several years. An Arabic CAPTCHA has recently been proposed to serve Arab users. Since there have been few scientific studies supporting a systematic design or tuning for users, this paper aims to analyze the Arabic text-based CAPTCHA at the parameter level by conducting an experimental study. Based on the results of this study, we propose an Arabic text-based CAPTCHA scheme with Fast Gradient Sign Method (FGSM) adversarial images. To evaluate the security of the proposed scheme, we ran four filter attacks, which led to a success rate of less than 5%. Thus, we developed a defensive method against adaptive attacks which is a standard for evaluating defenses to adversarial examples. This method is ensemble adversarial training, and it gave an accuracy result of 41.51%. For the usability aspect, we conducted an experimental study, and the results showed that it can be solved by humans in a few seconds with a success rate of 93.10%.