FIND THE FELINES
Three Microsoft researchers in Redmond have devised a better online test to determine who’s human and who isn’t.
Jeremy Elson, Jon Howell and John Douceur have turned a “pet project” into a more effective alternative to the twisted letters or distorted numbers that sites use to separate humans from Web robots or computer scripts trying to hack a site.
Those twisted numbers and distorted images are called “captchas” – and are used at sites before someone can buy a ticket, post a message or create an account.
“Captchas are used at Web e-mail sites like Hotmail, which can be attacked by robots that would try to sign up for thousands of e-mail accounts to send out spam,” said Elson.
The solution has worked adequately but has also created frustration, the researchers learned. Plus, sophisticated computers with character-recognition tools can do as well as people deciphering captchas.
Howell, Elson and Douceur latched onto a system that uses photos of cats and dogs. Instead of identifying letters, the Web site asks you to identify which of the dozen or so pet images are cats and which are dogs. If you get it right, the site knows you’re a person, said Elson.
The idea jumped into their brains when they were puzzling over how to help managers protect Web sites from bots or scripts that can get around captchas. One idea that caught their attention was developed by someone named “frozenbear,” who modified an idea from the once-popular Web site Hotornot.com.
Using images of men and women, frozenbear created a test requiring users to pick the people who were “hot,” then submit the answer. That little test is still being showcased at Hotcaptcha.com.
Douceur, Howell and Elson decided to use the same idea but with a large number of cat and dog photos, linked to the pet-adoption site Petfinder.com.
“Distinguishing dogs and cats is something that humans can do far better than computers,” said Elson.
It also improved on the subjective nature of selecting which of 10 young men or women qualified as attractive, a judgment call that’s often culturally based, said Douceur.
“Here was a very large database that really did have some ‘ground truth’ and which wasn’t open to cultural subjectivity,” said Douceur. After several months of development, they introduced the tool – called Asirra – at the recent TechFest showcase at Microsoft’s Redmond campus.
So far a number of sites are testing Asirra, which is free to use without a licensing fee. All Microsoft asks is that sites that display the Asirra tool add the “adopt me” buttons below each cat and dog photo. That sends the user to Petfinder where they can learn more about that animal.
Several blog sites and smaller Web sites are also testing the tool, as are a few much larger companies, said Douceur.
The social benefit will be the increased number of people who find pets they want to adopt.
The Web benefits could be increased security and higher consumer satisfaction. As many Web site managers have learned, customers often get frustrated when they fail a captcha test and then leave without completing a purchase.
“With a captcha there’s a trade-off between being too easy for a script to get around and being too hard or too annoying for a person to get around,” Douceur said. “We hope Asirra provides a level of security equal to (a captcha) without being annoying.”
One technical challenge for the researchers was getting access to 2.2 million cat and dog images from Petfinder. That large number – updated with 10,000 new images daily – ensures a high probability that no one can devise a hack to outsmart the human challenge, said Douceur.
“If I used just 15 or so pictures of cats and dogs, someone could spend a weekend putting them in a database and writing a program that would be able to identify which are cats and which aren’t,” Douceur said.
Because it’s a product from Microsoft’s lab, the three researchers post the code on their team Web site – at http://research. microsoft.com/asirra. Anyone can use it – even companies like Yahoo or Google, which are usually viewed as Microsoft competitors.