A novel CAPTCHA solver framework using deep skipping Convolutional Neural Networks.

Area of Science:

• Computer Science
• Cybersecurity
• Artificial Intelligence

Background:

• Completely Automated Public Turing Test to tell Computers and Humans Apart (CAPTCHA) systems are crucial for web security but vulnerable to Optical Character Recognition (OCR) attacks.
• Existing CAPTCHA-breaking methods, including deep learning, often lack robust validation and comprehensive feature extraction, necessitating improvements.

Purpose of the Study:

• To develop a more effective CAPTCHA solver using advanced deep learning techniques.
• To address the limitations of dataset-specific studies by creating a more generalizable CAPTCHA-breaking solution.

Main Methods:

• Utilized two public datasets of 4- and 5-character text-based CAPTCHA images.
• Developed and implemented a skip-connection-based Convolutional Neural Network (CNN) model.
• Employed a 5-fold cross-validation strategy, generating 10 distinct CNN models for evaluation.

Main Results:

• The proposed skip-connection-based CNN model demonstrated promising results in solving CAPTCHA codes.
• The models achieved superior performance compared to existing studies on the evaluated datasets.
• The 5-fold cross-validation provided robust validation for the CNN models' effectiveness.

Conclusions:

• The developed CAPTCHA solver shows significant potential in enhancing web security against automated attacks.
• The skip-connection-based CNN architecture offers a more confident and multi-aspect feature covering scheme for CAPTCHA recognition.
• This research contributes to the ongoing effort to improve CAPTCHA security against sophisticated computer vision-based threats.