Breaking google reCaptcha V2

摘要

reCaptcha is a service designed to protect websites from spam and abuse by telling humans and robots apart. In this research, we focused on maximizing accuracy on classifying road signs vs background, in order to break the road sign puzzles of Google's reCaptcha version 2. We trained two different convolutional neural networks, did extensive dataset collection, and developed a tool to visualize the results. Our transfer learning neural network based on GoogLeNet with 144 layers has a success rate of 94.25% for classifying road signs, while our self-built neural network with 24 layers has a success rate of 77.2%. Based on our results, we believe that we have successfully broken the road sign category in Google reCaptcha version 2.