Neural network CAPTCHA crackers

This paper describes several experiments using deep neural networks to break character-based image CAPTCHAs. The goal of our research was to see if one could develop a single neural network capable of breaking all character-based, image CAPTCHAs. Our main deep neural net uses convolutional neural network layers followed by a dense layer, and a recurrent recurrent neural network layer instead of the conventional method of CAPTCHA breaking based on segmenting and recognizing individual letters. Our experiments with these networks were conducted using a synthetically generated dataset of CAPTCHAs which is independently useful for future research. We trained on both fixed-and variable-length CAPTCHAs and our main neural net configuration was able to achieve accuracy levels of 99.8% and 81%, respectively.