Improving Robustness and Visibility of Adversarial CAPTCHA Using Low-Frequency Perturbation

Abstract

CAPTCHA is a type of Turing test used to distinguish between humans and computing machine. However, image-based CAPTCHAs are losing their function as Turing tests owing to the improvement of image recognition using machine learning. This paper proposes an Adversarial CAPTCHA that provides attacking resistance to CAPTCHAs by using Adversarial Example (AE) as well as maintaining visibility by reducing image degradation. The proposed CAPTCHA maintains the difficulty of solving CAPTCHAs using computing machine by adding resistance against the attack using a machine learning classifiers. The proposed CAPTCHA is evaluated using three evaluation experiments, i.e., the attack using a machine learning classifier, the image quality, and the solving workload. The three evaluation experiments show that an Adversarial CAPTCHA is resistant to the attack by machine learning and is as convenient as the existing CAPTCHA.