Completely Automated Public Turing test to Tell Computers and Humans Apart, more commonly referred to as CAPTCHA, is a test people often have to go through before completing a registration, making an online purchase, or submitting an application.
A CAPTCHA isn't really useful in and of itself, it's just the most efficient and effective way to prevent online algorithms or spambots from pretending to be humans and wreaking havoc.
But a team of researchers have successfully created an algorithm that can bypass CAPTCHAs, even complicated ones.
What Are CAPTCHAs?
Traditional CAPTCHAs are made up of jumbled letters oftentimes scattered with funky shapes, splashes of color, and other visual elements that distort and disrupt the letters-and-numbers combo. Their jaggedness has a purpose though: make it near impossible for robots to read, and possible for humans to understand, albeit with slight difficulty.
The ability to crack these CAPTCHAs, has become a crucial inflection point for artificial intelligence experts. Some have attempted and succeeded. Years ago, a company circumvented Ticketmaster's CAPTCHA and purchased a great number of concert tickets.
But such attempts were just responding to the weaknesses found in a certain type of CAPTCHA, which could easily be prevented with a few program alterations, says Dileep George, cofounder of AI startup Vicarious, which developed an entirely novel approach.
A New Model For Robots To Crack CAPTCHAs
Built on machine learning and loosely inspired by neuroscience, the Vicarious model is even more capable of cracking CAPTCHAs. As published on Thursday, Oct.26, in the Science journal, this new model parses text better compared with past models that have had less training, said George.
The problem with previous approaches is that they relied on something called deep learning to teach robots how to think like humans. Deep learning, a method in which layers are trained to respond in specific ways, only replicates some aspects of how the human brain functions, according to George. In deep learning, neurons need to be fed thousands upon thousands of images showing, say, a train, before it can begin to recognize what a train looks like.
Humans, meanwhile, don't actually need to see thousands, not even hundreds of pictures of trains to be able to recognize one.
"We found that there are assumptions the brain makes about the visual world that the [deep learning] neural networks are not making," said George.
Vicarious's approach is it builds internal models of the letters that the system is exposed to. This process, called a recursive cortical network, or RCN, works by teaching the system As and Bs and other different characters, after which it will try to build its own model of what those characters are supposed to look like. As a result, whenever it tries to look at a new image — a CAPTCHA full of confusing letters, numbers, and shapes — it compares that with previously seen ones and is able to determine, for example, that this A doesn't like the typical A shape because it's obstructed, or hidden behind another letter.
More Than Just Cracking CAPTCHAs
According to the paper, the Vicarious model successfully solved reCAPTCHAs at an accuracy rate of 66.6 percent. But this research has far more complex goals than just cracking CAPTCHAs — it's actually about teaching robots how to think more like humans.
"We are working on several tasks in robotics," said George. "Data efficiency and reasoning are very important when robots deal with unstructured environments."
