Adversarial Text-Based CAPTCHA Generation Method Utilizing Spatial Smoothing

The development of deep learning (DL) techniques has enabled to crack traditional text-based CAPTCHA (Com-pletely Automated Public Turing test to tell Computers and Humans Apart), which results in new security issues. As a coun-termeasure against DL based attacks, the adversarial CAPTCHA is well suited since it can increase the difficulty of machine recognition while ensuring human readability. However, spatial smoothing can negate the effectiveness of adversarial CAPTCHAs because adversarial noises in them are subject to averaging pixels. As far as we know, there are no effective counters against spatial smoothing, whereas it is the critical problem which facilitates spreading automated attacks. Therefore, to address the unsolved problem, in this paper, we propose an adversarial text-based CAPTCHA generation method utilizing spatial smoothing. We focus on the fact that when spatial smoothing is applied to an image, the amount of information it carries decreases, making the whole image blurred. Spatial smoothing is only viable as an attack when the mitigation of the adversarial noise has a larger impact than the whole image getting blurred. Thus, when the degree of spatial smoothing exceeds a certain threshold, the impact of the two aspects reverses, and the difficulty of the recognition increase. By utilizing this phenomenon in the generation of CAPTCHAs, the proposed method can indirectly neutralize the intended effect of spatial smoothing by attackers, preventing the recognition rate from increasing. Our evaluation shows the proposed method can reduce the recognition rate by up to 34%, compared to the conventional method. Besides, an experiment on human recognition rates marked 73.67%, showing that human recognition is maintained at an acceptable level.