A Generic Solver Combining Unsupervised Learning and Representation Learning for Breaking Text-Based Captchas

Although there are many alternative captcha schemes available, text-based captchas are still one of the most popular security mechanism to maintain Internet security and prevent malicious attacks, due to the user preferences and ease of design. Over the past decade, different methods of breaking captchas have been proposed, which helps captcha keep evolving and become more robust. However, these previous works generally require heavy expert involvement and gradually become ineffective with the introduction of new security features. This paper proposes a generic solver combining unsupervised learning and representation learning to automatically remove the noisy background of captchas and solve text-based captchas. We introduce a new training scheme for constructing mini-batches, which contain a large number of unlabeled hard examples, to improve the efficiency of representation learning. Unlike existing deep learning algorithms, our method requires significantly fewer labeled samples and surpasses the recognition performance of a fully-supervised model with the same network architecture. Moreover, extensive experiments show that the proposed method outperforms state-of-the-art by delivering a higher accuracy on various captcha schemes. We provide further discussions of potential applications of the proposed unified framework. We hope that our work can inspire the community to enhance the security of text-based captchas.