A Novel Deep Learning Based Fully Automated Framework for Captcha Security Vulnerability Checking

From breaching the service and allowing humans to surpass it upon correct verification, nowadays, it is achievable by bots and machines using machine learning, making the service vulnerable. It can pose several challenges like hacking, crashing, exploiting to name a few. In the present work, a novel framework namely BypassCaptcha model to test the security strength of web services via captchas is proposed. This model examines the vulnerability of the security via automation and captcha decoding using deep learning models particularly, i.e. Convolutional Neural Network, Recurrent Neural Network, and Connectionist Temporal Classification Loss. It involves full automation processing i.e. from opening the service, entering the credentials, getting a captcha from the service, and inputting the right decoded captcha. This complete process is dynamic. The credentials are provided via file, read during runtime, and are inputted at their required place. For training the model, a dataset combinely having four different types of captcha i.e. arc, dotted, rotated, and noisy is used. The automation process is working satisfactorily on specific services but it is still not a convenient way for a large number of services at a time. The proposed model is having a Val_Loss of 97% in the clear captcha case.