Multiview deep learning-based attack to break text-CAPTCHAs.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Text-based Completely Automated Public Turing Test To Tell Computer and Humans Apart (CAPTCHA) schemes are widely used to distinguish humans from bots.
• These schemes are popular due to their lower computational cost compared to other CAPTCHA types.
• However, their effectiveness against sophisticated automated attacks remains a concern.

Purpose of the Study:

• To investigate the security vulnerabilities of current state-of-the-art text-CAPTCHA schemes.
• To propose and evaluate a novel Multiview deep learning system for breaking text-CAPTCHA.
• To highlight the weaknesses inherent in existing text-CAPTCHA designs.

Main Methods:

• Developed a Multiview deep learning system combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs).
• The system extracts correlational features from multiple views to enhance generalization and classification accuracy.
• Preserves spatial and sequential information of the input text-CAPTCHA images.

Main Results:

• Achieved high average accuracies between 93.6% and 100% across eight diverse datasets.
• Successfully broke text-CAPTCHA schemes in an average time ranging from 0.0032 to 0.21 seconds.
• An ablation study confirmed the system's superiority over human users in solving text-CAPTCHA.

Conclusions:

• The proposed Multiview deep learning system significantly outperforms existing state-of-the-art methods, with an accuracy advantage of nearly 40%.
• Current text-CAPTCHA schemes exhibit critical weaknesses, as demonstrated by the system's ability to bypass them effectively.
• The findings necessitate a re-evaluation of text-CAPTCHA security and the development of more robust authentication mechanisms.