Vulnerable JavaScript functions detection using stacking of convolutional neural networks.

Area of Science:

• Computer Science
• Cybersecurity
• Software Engineering

Background:

• Web application security is critical to prevent cyberattacks.
• Static analysis is a common method for identifying software vulnerabilities.
• Existing methods for detecting vulnerable JavaScript code often suffer from low accuracy and high false positives/negatives.

Purpose of the Study:

• To propose a novel approach for identifying vulnerable JavaScript functions using an ensemble of convolutional neural networks (CNNs).
• To enhance the accuracy and reliability of detecting security vulnerabilities in web applications.
• To address the limitations of current static analysis techniques for JavaScript code.

Main Methods:

• Developing an ensemble of convolutional neural networks (CNNs) models.
• Utilizing vulnerable code information and code features for detection.
• Implementing a stacked CNN approach combined with sampling techniques (misbalancing, random undersampling, random oversampling).
• Training and evaluating the model on publicly available JavaScript code blocks.

Main Results:

• The stacked CNN approach demonstrated high effectiveness in detecting vulnerable JavaScript functions.
• The proposed method achieved an accuracy rate of approximately 98%.
• The approach showed robustness and usability for real-world security applications.

Conclusions:

• The ensemble CNN model offers a significant improvement over existing methods for identifying vulnerable JavaScript code.
• This research provides a more effective solution for protecting web applications from cyber threats.
• The findings contribute to a safer online environment by enhancing system security.