STACAD: A Large-Scale Empirical Study of Paired LLM-Paraphrase Detection and Attribution Using Transformation and Stylometric Features

Replication package for the paper STACAD: A Large-Scale Empirical Study of Paired LLM-Paraphrase Detection and Attribution Using Transformation and Stylometric Features. STACAD studies the paired problem of deciding whether a candidate source file is an LLM paraphrase of a human-written reference and, if so, which model produced it. The study spans seven LLMs (Gemini 3 Flash, GPT-5 Nano, Claude 3 Haiku, Qwen3 Coder 30B, DeepSeek V3, Grok 4 Fast, Devstral), seven programming languages (Python, C, C++, Java, C#, Go, PHP), and 103,768 paired samples drawn from 21,309 GitHub repositories. A weighted soft-voting ensemble of XGBoost, LightGBM and CatBoost over a 105-feature paired vocabulary (85 transformation features in 11 categories plus 20 stylometric measures) is evaluated under five-fold repository-level GroupKFold cross-validation. Binary paraphrase detection reaches 99.93% accuracy; seven-class attribution reaches 57.80% accuracy and 57.95% macro F1, ahead of fine-tuned CodeBERT and an attention classifier on the same features. Contents (extract stacad-replication-package.zip to STACAD/): the paired Task 1 (detection, 207,536 records) and Task 2 (7-class attribution, 103,768 records) corpora; a 5,000-sample adversarial-robustness pilot subset; precomputed feature matrices and per-fold result artifacts under cache/; a cache-aware end-to-end notebook (notebooks/final_STACAD.ipynb); the corpus-construction scripts; and an integration note for the third-party LPCodec baseline. The notebook loads the shipped artifacts and reproduces the headline numbers on CPU within a few minutes; the provided rebuild flags regenerate every result from the raw corpus. See README.md for full instructions and results_summary.json for the reported metrics. The code is released under the MIT License and the data under CC BY 4.0; bundled human source code retains the original permissive licenses (MIT/Apache/BSD) of the crawled repositories.