An End-to-End Attack on Text-based CAPTCHAs Based on Cycle-Consistent Generative Adversarial Network

摘要: As a widely deployed security scheme, text-based CAPTCHAs have become more and difficult to resist machine learning-based attacks. So far, many researchers conducted attacking research on by different companies (such as Microsoft, Amazon, Apple) achieved certain results.However, most of these attacks some shortcomings, such poor portability attack methods, requiring series data preprocessing steps, relying large amounts labeled CAPTCHAs. In this paper, we propose an efficient simple end-to-end method based cycle-consistent generative adversarial networks. Compared with previous studies, our greatly reduces the cost labeling. addition, has high portability. It can common CAPTCHA schemes only modifying few configuration parameters, which makes easier. Firstly, train synthesizers cycle-GAN generate fake samples. Basic recognizers convolutional recurrent neural network are trained data. Subsequently, active transfer learning is employed optimize basic recognizer utilizing tiny real-world Our approach efficiently cracked 10 popular websites, indicating that likely very general. Additionally, analyzed current anti-recognition mechanisms. The results show combination mechanisms improve CAPTCHA, but improvement limited. Conversely, generating complex may resources reduce availability