Neural CAPTCHA networks

To protect against attacks by malicious computer programs, many websites apply the CAPTCHA (short for completely automated public turing test to tell computers and humans apart) technique for security protection. The distortion, rotation and deformation of the characters or puzzles in CAPTCHAs increase the difficulty for machines to automatically recognize them. State-of-the-art CAPTCHA recognition algorithms generally use convolutional neural networks (CNNs) without considering the spatially sequential property of the characters/image features. To address this problem, we propose a new CAPTCHA recognition algorithm called neural CAPTCHA networks (NCNs). NCNs use a convolutional structure to extract CAPTCHA image features, and use bidirectional recurrent modules to learn the spatially sequential information in CAPTCHAs. We have applied NCNs to recognize text-based CAPTCHAs, including arithmetic operation, character recognition and character matching CAPTCHAs, and puzzle-based CAPTCHAs. For arithmetic operation and character recognition CAPTCHAs, we obtained 100% accuracy on the SOIEC CAPTCHA dataset, for the character matching task, we obtained 99.3% accuracy on the SOIEC CAPTCHA dataset, while for the puzzle-based CAPTCHAs, we obtained 98.13% accuracy. These experimental results demonstrate the advantages of NCNs over related state-of-the-art approaches for CAPTCHA recognition.

Duke Scholars

Author Kaizhu Huang DKU Faculty