The war against bots is never-ending, though hopefully it doesn’t end in the Skynet-type scenario we all secretly expect. In the meantime, it’s more about cutting down on spam, not knocking down hunter-killers. Still, the machines are getting smarter and simple facial recognition may not be enough to tell you’re a human. Machines can make faces now, it seems — but they’re not so good at answering questions with them.
Researchers at Georgia Tech are working on a CAPTCHA-type system that takes advantage of the fact that a human can quickly and convincingly answer just about any question, while even a state of the art facial animation and voice generation systems struggle to generate a response.
There are a variety of these types of human/robot differentiation tests out there, which do everything from test your ability to identify letters, animals and street signs to simply checking whether you’re already logged into some Google service. But ideally it’s something easy for humans and hard for computers.
It’s easy for people to have faces — in fact, it’s positively difficult to not have a face. Yet it’s a huge amount of work for a computer to render and modify a reasonably realistic face (we’re assuming the system isn’t fooled by JPEGs).
It’s also easy for a person to answer a simple question, especially if it’s pointless. Computers, however, will spin their wheels coming up with a plausible answer to something like, “do you prefer dogs or cats?” As humans, we understand there’s no right answer to this question (well, no universally accepted one anyway) and we can answer immediately. A computer will have to evaluate all kinds of things just to understand the question, and double-check its answer, then render a face saying it. That takes time.
The solution being pursued by Erkam Uzun, Wenke Lee and others at Georgia Tech leverages this. The prospective logger-in is put on camera — this is assuming people will allow the CAPTCHA to use it, which is a whole other issue — and presented with a question. Of course, there may be some secondary obfuscation — distorted letters and all that — but the content is key, keeping the answer simple enough for a human to answer quickly but still challenge a computer.
In tests, people answered within a second on average, while the very best computer efforts clocked in at six seconds at the very least, and often more. And that’s assuming the spammer has a high-powered facial rendering engine that knows what it needs to do. The verification system not only looks at the timing, but checks the voice and face against the user’s records.
“We looked at the problem knowing what the attackers would likely do,” explained Georgia Tech researcher Simon Pak Ho Chung. “Improving image quality is one possible response, but we wanted to create a whole new game.”
It’s obviously a much more involved system than the simple CAPTCHAs we encounter now and then on the web, but the research could lead to stronger login security on social networks and the like. With spammers and hackers gaining computing power and new capabilities by the day, we’ll probably need all the help we can get.