Increasing the Security of Video CAPTCHAs through Text Detection and Removal Type 2: Innovative Student Research Proposal
Dave Snyder, Chester F. Carlson, Richard Zanibbi · 2011
CAPTCHA tests are used to distinguish humans and machines online for the purpose of preventing automated abuses of services such as web mail accounts. Commonly CAPTCHAs require distorted text to be transcribed, but the distortion needed to protect against Optical Character Recognition (OCR) makes them dicult for human users. Previously the Faculty Sponsor and a former graduate student developed the Video CAPTCHA, where three words describing a video are submitted, and the test is passed if one of the submitted words matches a ‘valid’ word automatically generated from video tags in the database. Human and attack pass rates are comparable with other CAPTCHAs for attacks submitting words with the highest estimated prior probabilities, but augmenting this attack with OCR will produce unacceptably low security. In the proposed work, we seek to mitigate this risk through developing state-of-the-art algorithms for detecting text in video and using inpainting to unobtrusively replace detected text regions. Our goals are to increase the security of the CAPTCHA against such OCR-based attacks while maintaining human pass rates, and to rene our Video CAPTCHA so that it may be used in practice.