CAPTCHAs Are Still in Danger: An Efficient Scheme to Bypass Adversarial CAPTCHAs

Dongbin Na, Namgyu Park, Sangwoo Ji, Jong Kim

Abstract: Completely Automated Public Turing test to tell Computers and Humans Apart (CAPTCHA) is a challenge that is used to distinguish between humans and robots. However, attackers bypass the CAPTCHA schemes using deep learning (DL) based solvers. To defeat the attackers, CAPTCHA defense methods utilizing adversarial examples that are known for fooling deep learning models have been proposed. In this paper, we propose an efficient CAPTCHA solver that periodically retrain the solver model when its accuracy drops. The proposed method uses incremental learning that requires a small amount of data while achieving high accuracy. We demonstrate that the proposed solver bypasses the existing defense methods based on a text-based CAPTCHA scheme and an image-based CAPTCHA scheme.