Computer scientists have developed software that easily defeats audio CAPTCHAs offered on account registration pages of a half-dozen popular websites by exploiting inherent weaknesses in the automated tests designed to prevent fraud.
Decaptcha is a two-phase audio-CAPTCHA solver that correctly breaks the puzzles with a 41-percent to 89-percent success rate on sites including eBay, Yahoo, Digg, Authorize.net, and Microsoft's Live.com. The program works by removing background noise from the audio files, allowing only the spoken characters needed to complete the test to remain.
In virtually all of the tests, Decaptcha was able to correctly solve the puzzle at least once in every 100 attempts, making the technique suitable for botmasters with large armies of compromised computers. The high success rate was largely the result of the ease in removing sound distortions known as background noise, intermediate noise, and constant noise inserted into the background to throw off speech-recognition programs. Most audio-based CAPTHA systems are wide open to the attack with the notable exception of the Google-owned Recaptcha.net, which uses a different approach known as semantic noise.
"Our results indicate that non-continuous audio captcha schemes built using current methods (without semantic noise) are inherently insecure," the scientists wrote in a recently published research paper. "As a result, we suspect that it may not be possible to design secure audio captchas that are usable by humans using current methods. It is therefore important to explore alternative approaches."
Decaptcha uses a supervised algorithm that must be trained for each CAPTCHA scheme being targeted. Training requires feeding a set of puzzles with their answers into the program. Eventually, Decaptcha was able to identify the sound shapes in the underlying audio file by comparing them to a large sample of sounds already cataloged. The researchers generated 4.2 million audio CAPTCHAs.
The paper is only the latest reminder of the flaws in CAPTCHAs, which are designed to prevent scripts from registering email accounts, and carrying out other automated attacks, by presenting the user with a problem that's hard for computers to solve. Real-world attacks against audio-CAPTCHAs from Microsoft have already been used by the Pushdo spam botnet to create fraudulent email accounts on Live.com. More traditional CAPTCHAs, which require a user to recognize a word buried in a distorted image, have been successfully defeated for years, with one of the more recent examples being an optical character recognition attack on Google.
After attacks come to light, website operators typically make changes that block specific technique. Researchers then revise their attacks, requiring more changes to be made in the targeted CAPTCHA schemes.
The latest research suggests web developers may have to make permanent changes to the audio CAPTCHAs, which are offered for visually impaired users.
"Our experiments with commercial and synthetic captchas indicate that the present methodology for building audio captchas may not be rectifiable," they wrote. "Besides Recaptcha, all of the commercial schemes we tested used combinations of constant and regular noise as distortions. 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."
The paper was authored by Elie Bursztein, Hristo Paskov, and John Mitchell of Stanford University, Romain Beauxis of Tulane University, Daniele Perito of INRIA and Celine Fabry. A PDF of the report is here. ®