Mitigating Deep learning Attacks Against Text Image CAPTCHA Using Arabic Scheme

Abstract

Confidentiality and availability are the main concerns of website stakeholders, entailing service-security trade-offs. Many solutions seek to improve security. The most popular one is the completely automated public Turing test (CAPTCHA) security framework, which distinguishes between human and bot activity by asking the user to perform a certain task, such as solving a mathematical equation or retyping text written in an image. Many people believe that websites using CAPTCHAs are secure, but today’s attackers using innovative technology, such as deep learning, that can break them in milliseconds. Many solutions have been proposed to overcome this issue, mainly relying on increased CAPTCHA complexity or technical CAPTCHAs too difficult for normal users. In this case, many customers may go to websites that use simple CAPTCHAs, at the expense of websites with more complex ones. This paper proposes a robust and simple image-based CAPTCHA using an Arabic scheme and noise to make breaking this captcha very difficult, even for deep learning approaches. The proposed model is evaluated using two-factor, which focuses on the usability and the ability to beak the proposed CAPTCHA. The first test has been done via an empirical experiment involving 16 people from different sectors, and the second test has involved breaking the model using Convolutional Neural Network (CNN). The experimental results demonstrate the superiority of the proposed model in security and usability.