言語的作問技法を用いたCAPTCHAの構築 = Design and implementation of CAPTCHA using verbal test

TLDR: A new CAPTCHA system which can differentiate between humans and software agents pretending to be humans, based on their different contextual cognition, is proposed, which works without relying on the specific perceptual abilities of the users.

Abstract: CAPTCHA (Completely Automated Public Turing test to tell Computers and Human Apart) is now the de facto standard security technology to protect on-line registration systems from malicious software. CAPTCHA systems generate several kinds of AI (Artificial Intelligence) problems which are di fficult for software agents but easy for humans. A big social problem is that most visually-impaired people cannot pass current CAPTCHAs. Most conventional CAPTCHAs employ AI problems requiring visual recognition, so they are not accessible to visually-impaired people. Audio CAPTCHAs are an alternative to visual ones, but several researchers have pointed out that state-of-the-art audio ones are too di fficult for visually impaired people. In this paper, we propose a new CAPTCHA system, which generates tests in verbal style. Our CAPTCHA system can differentiate between humans and software agents pretending to be humans, based on their different contextual cognition. It therefore works without relying on the specific perceptual abilities of the users. In our test, we utilize open documents for material of the tests. Note that there is quite a large amount of documents on the web, so we can generate brand-new tests every time. This is di fferent from conventional studies. One criticism is that adversaries can look for the phrases of the tests from the Internet and obtain several hints. Our system hides the sources by substituting the consonants of the phrases against such adversaries. The mechanism is similar to the phenomenon of “consonant gradation” in natural languages. The substitutes make it harder for adversaries to look for the sources because they have difficulty finding the original phrases from the erroneous ones. We apply our idea to three kinds of verbal tests: (a) M rkov-chain Phrase Test , which involves distinguishing between natural and machine-synthesized phrases, (b) Machine-translated Phrase Test, which involves distinguishing between natural and machine-translated phrases, and (c) Topic Detection Test, which is a choice of the common topic from several short-texts. We then implement them as CAPTCHA programs, and evaluate their performances as follows: (1) ability to be used for a Turing test, which must be easy for humans to solve, (2) ability to generate new tests without limitation on amount, and (3) ability to hide the sources of the phrases which appear in the test. Consequently, we have clarified the feasibility of our proposal as CAPTCHA for the visuallyimpaired.