Automatic recognition of captchas using neural networks

Abstract

The Completely Automated Public Turing test to tell Computers and Humans Apart (CAPTCHA) is a test used to differentiate between humans and machines. In this thesis, we have implemented a deep neural network architecture to automatically recognize CAPTCHAs. The main idea behind this thesis is to test the reliability of test-based CAPTCHAs in differentiating between humans and machines. We utilized convolutional neural networks instead of the conventional strategies of CAPTCHA breaking which, typically, use segmentation techniques. Our model is comprised of 8 layers to learn textbased images; five of which are used for convolution and the remaining three are dense feedforward layers. We attempted various parameters to our model, including dropout rate, maxpooling value and the number of samples used. We generated a CAPTCHA dataset of two different types; colored and gray scaled images CAPTCHAs and had the capacity to get accuracy levels of 98.6% and 100%, respectively.