If you spend any amount of time on the Internet, you have certainly run across CAPTCHA on a few occasions. CAPTCHA is the security device that forces you to read squiggly words and type them into the box to prove you are a person and not a spamming robot sent from the future to troll forums and steal all of our memes. The schtick for CAPTCHA is that computer programs can't read the words, but that is no longer true.
AI has been created that is able to defeat CAPTCHA security meaning all our memes aren't safe after all. The AI was inspired by a human brain, specifically the human visual cortex. To rub salt in CAPTCHA's wound, the new AI needed 50,000 times fewer training examples to crush the security than other state-of-the-art programs.
"We just have to accept that, as computer vision systems improve, traditional text-based CAPTCHA systems no longer offer the protection they used to," said study co-author Miguel Lazaro Gredilla of Vicarious AI.
CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) was originally meant to keep spamming bots at bay and ensure that only humans could read and type in the words. Today, machine learning and more raw computation power
To keep ahead of the evil bots, CAPTCHA had become more and more complicated. Most people who have used CAPTCHA have probably gotten the words wrong at least once. Apparently, humans can only solve Google's more complex reCAPTCHAs 87% of the time.
Vicarious AI is the company that created the new AI and since it is based on neuroscience, the AI spent much less time on training. The method that the scientists used is a focus on "contour continuity" which is the way that the human brain is able to distinguish between edges of an object even when that object is partially blocked by another object.
The scientists call their invention a "recursive cortical network" and say that it is much simpler than a neural network with the recursive cortical network having only a few processing layers. The first step in that processing path is to analyze pixels to figure out if they form part of an edge of an object. The data is then passed through the rest of the network and as it works through the layers, the contours of an image are formed. The final layer of the recursive cortical network is able to produce the object in question, meaning figure out it is a letter "A".
The fancy AI network is able to work even if CAPTCHA letters overlap. If a machine is able to defeat a CAPTCHA more than 1% of the time, it is considered to be useless. Vicarious AI's fancy cortical network AI was able to beat Google's reCAPTCHA 67% of the time after a mere five training examples per character. Comparative neural networks needed 50,000 times more training examples. The CAPTCHAs that Yahoo and Google
The next generation of CAPTCHA is already being used including one Google has called "invisible CAPTCHA" that forces users to click a button before the CAPTCHA is displayed to answer. The idea here is that the human imperfections are noted as the mouse is moved and buttons clicked. "From the moment the webpage loads to the click, you're doing things," Dr. Adnene Guabtni, a cybersecurity researcher at CSIRO's Data61 said.
"If the button is to the left of your mouse, you move the cursor. But when you do that, you will wobble a little. That is detected. "It knows you're human because you have imperfections. A machine would try to go in a straight line."
Dr. Guabtni says it is only a matter of time until machines can crack this sort of CAPTCHA as well, meaning we are not safe from the machines. AI is becoming more flexible and powerful, and can beat humans at some of the most complicated games in history. Google's DeepMind Alpha Go Zero defeated the best human Go players.