Akrout and his fellow researchers fool the one section of reCAPTCHA using a type of machine learning called reinforcement learning, in which a software agent tries to find the best possible path, encouraged via rewards for each step it takes in the right direction. Their system lays a grid of squares across the page, working the mouse diagonally through the grid towards the “I am not a robot button”. Given positive reinforcement if it’s successful and negative if not, the system learns the right ways to move to convince reCAPTCHA the mouse is being controlled. The paper claims it works with 97.4 per cent accuracy. At the time of publication, Google had not responded to a request for comment about the paper.
Read More.. Source Wired