Apps optimize

Asirra: CAPTCHAs with a heart

Discover why Microsoft Research's Asirra (Animal Species Image Recognition for Restricting Access) project makes blogger Justin James happy.

Every once in a while, I come across a project or piece of software that really puts a smile on my face. About a month ago, I found Asirra (Animal Species Image Recognition for Restricting Access) from Microsoft Research. It is a CAPTCHA system that asks the user to decide which photos are dogs and which photos are cats.

Why does Asirra make me happy? Well, it does some pretty neat things. It brings a significantly stronger CAPTCHA methodology into play, while simultaneously making it much easier for humans to use. (By stronger, I mean that it is less likely that software can fake it out.) I also like the connection Asirra has with Petfinder.com: The photos are from Petfinder's database of animals, which have already been identified as dogs or cats by the animal shelters. This provides a database of more than three million photos to draw from. Best of all, each image has an Adopt Me link that goes to the Petfinder site. You can see an example of Asirra in action below.

Figure A

The "strength" of the photo CAPTCHA compared to text is pretty clear to me. Computers are still really, really bad at any kind of image recognition, particularly when it is trying to match an image to a concept instead of a known image. For example, computers are just now able to compare two photos of the same person and identify them as the same person even if the clothing, facial hair, glasses, etc. have changed. It's still not possible to match a photo of a cat with the word cat with any kind of reliability. Compared to the typical text CAPTCHAs, which are easily hacked, photos are definitely stronger.

In terms of usability, Asirra is an improvement over the text CAPTCHAs. CAPTCHAs still have usability problems, such as not working for blind users and being a fairly confusing and workflow breaking part of a process. After all, signing up for an e-mail account or trying to post a message does not involve "trying to read some squiggly text" or "identifying photos of cats" as part of the process that a business analyst would draft. That said, I believe that the photo CAPTCHAs are a bit less difficult than text CAPTCHAs, since instead of struggling to read something that is intentionally obscured, you are simply making a binary decision based on a clear image.

Finally, the Adopt Me link is nice icing on this cake. It is always great to see a project that can do a bit of good in the world, and this one does it in a way that I am not used to seeing in the IT industry. It is fairly easy for me to feel a little jaded by the dozens of open source projects that promise to bring peace and harmony to the world by freeing the bits, or to be a touch sarcastic when a megacorporation donates a few leftover PCs to some poor African village that lacks running water. I hear a lot of rah-rah about "giving back to the community" but, frankly, it is tough to do so when you work 50+ hour workweeks and have a family waiting for you at home. So when I learn about a project or software that is free and has the chance to do some good in the process, I get a bit warm and gooey on the inside.

If you have a need for a CAPTCHA and are not set on one in particular, please give Asirra a chance. You'll be getting a more usable, less hackable system, while helping a dog or cat find a good home.

J.Ja

About

Justin James is the Lead Architect for Conigent.

8 comments
le3293
le3293

And Totally unusable for persons that are blind or have low vision. Leading to your web owner potentially facing legal liability for discrimination against persons with disabilities! Just pictures is a no go!

GTGeek88
GTGeek88

Hmmm...maybe I should work for Microsoft research. After I heard about CAPTCHAs being hacked, I figured we simply needed a different, stronger CAPTCHA and thought about animal identification with some of the features I read about here. Glad to see someone's thinking this way.

shirtbird
shirtbird

What prevents a program from simply guessing "cat" or "dog" - you've got a 50/50 chance, so will there be additional animals to try and make this a little more difficult? Or am I missing something?

Justin James
Justin James

Yup, that is a major drawback of *any* CAPTCHA system. They usually have a link for vision impaired users to follow, which takes them through a different route, to get around this. J.Ja

Justin James
Justin James

Indeed, MSR cites KittenAuth on their site (http://research.microsoft.com/asirra/). The big difference is that Asirra has a much larger database (over 3 million photos) than similar sites, because of the connection to Petfinder.com. In this case, I suspect that the "research" end of this for them is not to make a stronger CAPTCHA in and of itself, but to examine how people work with these things in general. J.Ja

Justin James
Justin James

In general, Microsoft Research looks like a really exciting place, in terms of the projects that they are working on. I could write about nothing but stuff I have found on their Web site, and have material for the next year. It is unfortunate that so few companies put the money into basic R&D. AT&T used to do this (Bell Labs, which became Lucent), I think IBM *might* still do this. For a while, Microsoft & IBM led the nation in patent apps filed per year, which is a good measure of their R&D efforts, and this is probably still the case. GE used to (probably still does) a lot of this stuff too. But in terms of computers, Microsoft and maybe IBM are the only ones doing largescale research into far too many topics. This is a topic for another time (and it's on my list of things I want to write about in the near future), but I think that if people want to see a less Microsoft-centric future, there needs to be more groups (not necessarily companies) doing this exsact kind of work. J.Ja

Justin James
Justin James

I didn't write it out, but I had a screenshot of it (which didn't make it into the final post) that made it more clear why you can't just guess. You are given something like 10 pictures and need to select the 5 or so cats in it. While it is still "guessable", it is pretty darned unlikely to get it right like that. I would also imagine that the service sees someone pounding away doing random guesses & flags them, although that is not stated for sure. J.Ja