Execution-Guided Detection of Semantic Code Perturbations

This paper proposes execution-guided detection: a framework that runs AI-generated code in a sandboxed environment, captures execution traces across diverse test inputs, and classifies code as clean or adversarially perturbed based on behavioral differences. Key findings:- Boundary inversion detection improves from 0% (static analysis) to 87.5% (execution traces)- Import aliasing detection improves from 7% to 58.2%- Hybrid detector (execution + AST features) achieves 0.904 accuracy and 0.942 F1 score- Wilcoxon signed-rank test confirms significance (p=0.0074) This is Paper 3 in the AI Code Security research series:- Paper 1: Detecting Adversarial Perturbations (DOI: 10.5281/zenodo.22349074)- Paper 2: Adversarial Code Review (DOI: 10.5281/zenodo.22357507)- Paper 3: Execution-Guided Detection (this work) Code: https://github.com/amanmukati09/execution-guided-detection