You're Doing it Wrong

Just a funny thing I came across while desperately trying to avoid working on some busted ass queries: Kitty authentication.

What is Kitty Authentication? Its a form of captcha that, instead of using funked text, uses images of cats and dogs and asks the user to discriminate between the two. Its an example of complex pattern matching, something that is easy for humans and horribly difficult for computer programs.

The idea is that the limitations of standard captchas ensure that they will be useless within the next few years. These limitations are nonexistant when asking users to match complex patterns. Current captchas suffer primarily from two limitations. First, there is a limit as to how obscured an image can be. In order to make text-based captcha work, you have to severely beat the fuck outta it. However, if you maul the text too much, people can't figure out what the text says. Because of this, it is pretty much guaranteed that OCR technologies will eventually catch and surpass humans in identifying obscured text. That's because OCR technologies are geared twoards distinguishing text within a badly scanned image. Second, the number of possible characters that make up a captcha is relatively low (most captchas only use the 26 letters of the alphabet; some add 2-9 as well). This means that there is a very limited set of possibilities of which a captcha character must be a member, thus reducing the complexity of identifying what a captcha is. Break a captcha image down into individual letters and it almost becomes a trivial task to crack.

By asking people to differentiate between two different, yet very similar, types of items, both of these limitations are avoided. Firstly, there is no need to obscure the image, making it easy for humans to identify. It is very simple for humans to look at a picture of a cat and a dog and tell the difference; it is virtually impossible, however, for a computer to do this reliably. Second, instead of there being a limited set of items to choose from, there is almost no theoretical limit to the number of pictures you can choose from. You could, like in the examples here, ask users to choose the difference between cats and dogs. Or, you could ask people to choose vegetables from among pictures of fruit. The choices are virtually limitless.

The single real limitation to this type of authentication is that there is a practical limit to the number of pictures you can use. This means computers could be trained to identify each image and what that image holds. The simple way around this would be to obfuscate the image. But doing that brings you back to the problem with captchas where you end up making them unidentifiable to humans.

In order to get around this issue, a group at Microsoft Research has partnered with petfinder.com in order to supply the images. Since Pet Finder has a gigantic (over a million) database of images of cats and dogs that are constantly being replaced, you avoid the issue of computers remembering a particular image and learning what it contains. Its a sweet idea. Not only do you get your captcha, but Pet Finder also gets free advertising. Each picture has a link to that animal's page on their site. So while you're answering a captcha challenge, you might also be adopting a pet!

Via Terry Zink's Anti-Spam blarg.

Programming Post by: McGurk at 03:28 PM | 1  Reply | Reply
Kick this post: