VDCRL: vulnerability detection with supervised contrastive code representation learning.

Area of Science:

• Computer Science
• Software Engineering
• Cybersecurity

Background:

• Deep learning models for code vulnerability detection show promise but struggle with generalization to new datasets.
• Existing methods often overfit to training data, leading to significant performance degradation on unseen code.
• Improved generalization is crucial for effective real-world application of automated vulnerability detection.

Purpose of the Study:

• To propose VDCRL, a novel framework for code vulnerability detection.
• To enhance the generalization ability of deep learning models in identifying software vulnerabilities.
• To achieve superior detection performance on both synthetic and real-world datasets.

Main Methods:

• VDCRL employs supervised contrastive code representation learning.
• Input-based and feature space-based data augmentation generate diverse code samples.
• A feature fusion encoder (SAFE) integrates source code and assembly instruction features.
• A Bidirectional Gated Recurrent Unit (BGRU) model is utilized for vulnerability detection.

Main Results:

• VDCRL demonstrated significantly improved generalization capabilities compared to state-of-the-art methods.
• The framework achieved superior vulnerability detection performance on real-world datasets.
• Feature fusion and supervised contrastive learning were key contributors to the enhanced performance.

Conclusions:

• VDCRL offers a robust solution for code vulnerability detection with enhanced generalization.
• The proposed framework effectively addresses the limitations of current deep learning approaches.
• VDCRL represents a significant advancement in ensuring software security through improved vulnerability detection.