This research introduces a novel approach to improving language model-based defect detection by applying semantic-preserving code transformations during model application.

Key Findings:

• Code mutations that preserve semantics can significantly improve defect detection capabilities
• The approach leverages metamorphic testing principles to enhance model performance without additional training
• Particularly effective for security vulnerability detection where code variations should maintain consistent predictions
• Demonstrates practical applications for improving existing security tools with minimal overhead

For security teams, this research offers a practical way to enhance vulnerability detection tools by testing equivalent code versions, leading to more robust security scanning without requiring model retraining.

Semantic-Preserving Transformations as Mutation Operators: A Study on Their Effectiveness in Defect Detection