A group of researchers from a company called Vicarious has developed a new artificial intelligence engine that does something that until recently seemed especially complex: to solve the puzzles posed by the CAPTCHA and reCAPTCHA systems to validate that users using these systems are really human beings.
The secret of this engine has been to study how the visual cortex works in humans to replicate it through the so-called Recursive Cortical Network (RCN). The artificial intelligence system is able to recognize the contours of the letters of the reCAPTCHA and can also deal with fuzzy or rotated letters. The security of the CAPTCHA and reCAPTCHA is therefore in question.
More precise than the human being
The way in which our visual cortex works makes it easy to recognize objects and shapes even if they are poorly oriented or the image is partially obscured or blurred. The RCN of Vicarious has a contour identification model that allows defining the shape of objects, and when certain common properties are recognized, these elements are grouped based on their physical proximity.
From there, relationships are established with which a kind of hierarchy tree is created, which in the end allows us to determine that those global contours that have been formed belong (or can belong) to a certain object. And, in this case, to a certain character, letter or number, which is what this AI system can detect with great success.
So much, in fact, that the Vicarious system is even better than the human being when it comes to recognizing that reCAPTCHA. Even we fail occasionally, and the average precision of the human being is 87%. The Vicarious system reaches 94% in some cases. As with DeepMind and its system for playing Go perfectly, Vicarious’s model is very specific and very focused on a specific problem, but once again it shows how far these engines can go.
The achievement, however, is not designed to overcome the newly invisible reCAPTCHA that Google began to use in 2016. This system precisely studies the interaction with the website in which these verification methods are used: how we move the mouse or how long it takes in clicking on certain sites. That Google system (different from “I’m not a robot” that is also seen often) seems to be safe … for now.