Detecting generated code through style
TraceForest explores code provenance through stylometric and abstract syntax tree features. I developed an explainable random-forest classifier that distinguishes AI-generated code from human-written code with 94% accuracy.
Instead of returning only a label, the system uses SHAP explanations to identify which features influenced an individual prediction. This makes the result more useful in review and governance settings where a model’s reasoning matters.
From model to developer workflow
I built Python and Java adapters that return classification probabilities for automated pull-request review workflows. I also extended CodeGPTSensor through encoder fine-tuning with stylometric features, raising accuracy beyond 95%.
The model was packaged as a VS Code extension and has received more than 1,500 downloads, turning an experimental classifier into a tool developers can use in their normal workflow.