Breaking CAPTCHA characters using Multi-task Learning CNN and SVM

CAPTCHAs have become the canonical methods to distinguish humans from bots or malicious programs. Most of the CAPTCHAs deployed on the websites are text-based. They are completely reliant on the assumption that bots cannot read text from an image. However, in recent times, people have been able to come up with programs that are able to recognize CAPTCHAs with significant accuracy. This paper aims to present similar methods, which aids in breaking of security. The paper presents Multi-task Learning CNN (MTL-CNN) and SVM can independently recognize text-based CAPTCHAs. With MTLCNN, no pre-processing of image is required and still, it achieves promising results. However, SVM demands some manual labor to achieve the same.