Breaking CAPTCHAs on the Dark Web

On the Dark Web, several websites inhibit automated scraping attempts by employing CAPTCHAs. Scraping important content from a website is possible if these CAPTCHAs are solved by a web scraper. For this purpose, a Machine Learning tool is used, TensorFlow and an Optical Character Recognition tool, Tesseract to solve simple CAPTCHAs. Two sets of CATPCHAs, which are also used on some Dark Web websites, were generated for testing purposes. Tesseract achieved a success rate of 27.6% and 13.7% for set 1 and 2, respectively. A total of three models were created for TensorFlow. One model per set of CAPTCHAs and one model with the two sets mixed together. TensorFlow achieved a success rate of 94.6%, 99.7%, and 70.1% for the first, second, and mixed set, respectively. The initial investment to train TensorFlow can take up to two days to train for a single type of CAPTCHA, depending on implementation efficiency and hardware. The CAPTCHA images, including the answers, are also a requirement for training TensorFlow. Whereas Tesseract can be used on-demand without need for prior training.