Using automatic speech recognition for attacking acoustic CAPTCHAs: the trade-off between usability and security

Abstract

A common method to prevent automated abuses of Internet services is utilizing challenge-response tests that distinguish human users from machines. These tests are known as CAPTCHAs (Completely Automated Public Turing Tests to Tell Computers and Humans Apart) and should represent a task that is easy to solve for humans, but difficult for fraudulent programs. To enable access for visually impaired people, an acoustic CAPTCHA is typically provided in addition to the better-known visual CAPTCHAs. Recent security studies show that most acoustic CAPTCHAs, albeit difficult to solve for humans, can be broken via machine learning. In this work, we suggest using speech recognition rather than generic classification methods for better analyzing the security of acoustic CAPTCHAs. We show that our attack based on an automatic speech recognition system can successfully defeat reCAPTCHA with a significantly higher success rate than reported in previous studies. A major difficulty in designing CAPTCHAs arises from the trade-off between human usability and robustness against automated attacks. We present and analyze an alternative CAPTCHA design that exploits specific capabilities of the human auditory system, i.e., auditory streaming and tolerance to reverberation. Since state-of-the-art speech recognition technology still does not provide these capabilities, the resulting CAPTCHA is hard to solve automatically. A detailed analysis of the proposed CAPTCHA shows a far better trade-off between usability and security than the current quasi-standard approach of reCAPTCHA.