Strategic Security Survey: Global Threat, LocalPain

All current audio Captchas have an inherent flaw that leaves them susceptible to automated attacks by botnets or spammers.

That warning comes from a group of researchers from Stanford University, Tulane University, and INRIA in France, who created a program called Decaptcha, which they designed to defeat audio Captchas that are based on non-continuous speech.

A Captcha refers to any challenge-response technique that's used to assess whether a website is interacting with a person, or a computer. "To avoid imposing undue user friction, Captchas must be easy for humans and difficult for machines. However, the scientific basis for successful Captcha design is still emerging," said the researchers.

Typically, a Captcha involves warping text or adding lines to the text, then asking the user to correctly deduce the actual letters. Unfortunately, fooling the optical character recognition software used in scripted attacks appears to be difficult, since many Captchas are so distorted as to be unusable.

As far as website design goes, a visual Captcha is also a virtual brick wall for people with visual impairments. As an accessibility document from the World Wide Web Consortium notes, visual verification routines pose "a major problem to users who are blind, have low vision, or have a learning disability such as dyslexia." Accordingly, many websites offer an alternative form of challenge and response. Legally, some organizations, including government agencies in many countries, must offer an alternative.

One popular alternative is the audio Captcha, which as a challenge and response requires that a user key in a short sequence of characters that they hear spoken. To help fool attackers, some audio Captchas also distort the audio.

But in a recently published paper, the aforementioned researchers said that Decaptcha successfully defeated most commercially available audio Captchas, including Authorize (89% of the time), eBay (82%), Microsoft's Live.com website (49%), Yahoo (45%), and Digg (41%). Only 20 minutes of training--having a user manually input 300 Captchas to train Decaptcha--was required on each system to achieve the above results.

That's a problem, because it means that attackers can design, at low cost, scripts that bypass visual Captchas to instead target--and regularly defeat--audio Captchas, which would allow them to register large numbers of fake users, harvest usernames, or send spam.

Interestingly, one audio Captcha service, Recaptcha, was only defeated 1.5% of the time, which they attribute to the service using "semantic vocal noise." But that security protection comes at a price, they note, since 40% of the time, it also fooled actual users.

Furthermore, based on tests against sites that use audio distortions, computers may actually have the edge. "All in all, computers may actually be more resilient than humans to constant and regular noise so any schemes that rely on these distortions will be inherently insecure," they said.

"As a result, we suspect that it may not be possible to design secure audio captchas that are usable by humans using current methods," said the researchers. "Decaptcha's performance ... indicates that such speech Captchas are inherently weak and, because of the importance of audio for various classes of users, alternative audio Captchas must be developed."

In the new, all-digital InformationWeek supplement: Our 2011 Strategic Security Survey confronts the five biggest problems faced by midsize companies. Download it now. (Free registration required.)