Evaluation of Ergonomically Designed CAPTCHAs using Deep Learning Technology

• Published in: Journal of Information Processing Vol. 26; pp. 625 - 636
• Main Authors: Azakami, Tomoka, Shibata, Chihiro, Uda, Ryuya
• Format: Journal Article
• Language: English
• Published: Information Processing Society of Japan 2018
• Subjects: amodal completion; CAPTCHA; convolutional neural networks; ergonomic effects
• ISSN: 1882-6652; 1882-6652
• DOI: 10.2197/ipsjjip.26.625

Although most existing text-based CAPTCHAs use distorted images of alphanumerics, they have been criticized because large image distortions make it difficult for human beings to recognize the characters, despite the ease with which computers can eliminate distortions and consequently recognize them. Ergonomically designed CAPTCHAs, which exploit human-specific phenomena, are a solution to this problem. Ergonomic design enables humans to momentarily recognize them, while significantly increasing computational costs for machines to recognize characters. Recently, owing to the development of deep learning, the image recognition capability of machines has improved dramatically. Characters are easily recognized within reasonable time by deep convolutional neural networks (DCNNs), which have similar architectures to visual perception mechanisms of the brain. In this paper, to clarify whether ergonomically designed CAPTCHAs can withstand state-of-the-art methods of deep learning, we use several kinds of DCNNs to measure the classification rates of the characters displayed. With respect to ergonomic designs, we first use Amodal CAPTCHA proposed by Mori et al., which exploits the two human-specific phenomena of amodal completion and aftereffects. We secondly modify Amodal CAPTCHA by adding jagged lines to the edges of characters, aiming to prevent DCNNs from recognizing them correctly, since edges are one of the most fundamental features for DCNNs. Experimental results, however, show that both naive and jagged-lined Amodal CAPTCHAs are almost completely broken. Another approach we conducted is to use only complete characters without shielding as training data, assuming that attackers have no information about how amodal completion and jagged edges were applied. However, even for this assumption, the classification rate of DCNNs is still sufficiently high. On the whole, our results in this paper show that any ergonomic effects such as amodal completion and jagged edges are no longer countermeasures against character recognition by DCNNs.