Students: Teslin Roys

Faculty Advisor or Community Project Lead: Dr. Saif al Zahir


In this paper, we argue that the notion of a CAPTCHA (Completely Automated Public Turing Test to tell Computers and Humans Apart) system which sorts humans from machines with perfect accuracy is not merely technically challenging but tightly connected to the philosophy of AI and therefore fraught. We embrace Von Ahn et. al. [2003]’s suggestion that new CAPTCHA designs be based upon difficult artificial intelligence problems, but we further recommend that new designs target areas of lesser development absent the expectation that an optimal solution will be found. CAPTCHA systems are typically primarily visual, characteristically challenging a user to identify a string of alphanumeric characters in an image. Many CAPTCHA systems in everyday practical use specifically employ the optical text recognition (finding characters in an image) approach, but we observe that this approach is not robust against increasingly well-developed attacks in the literature. We review newer methods, based on either image recognition or language semantics. We design a new CAPTCHA system in the latter category, but unlike existing examples of either type our system is not based upon a static, manually collected database and is not susceptible to rudimentary random guess attacks. Preliminary testing shows near-perfect acceptance of human users by the system, where comparable alternatives in the image recognition family have more inconsistent results. Using a dynamic social feedback mechanism where users score new phrases, this system is capable of identifying new synonyms and could also offer a measure of quality for automatically generated text.



Comments are closed