Machine Learning-Based Intrusion Detection for ROS2 Cyber-Physical Systems: An Evaluation of Random and Source-File-Aware Data Splitting

This technical report presents an independent research study on machine-learning-based intrusion detection for ROS2-based Cyber-Physical Systems (CPS). The study evaluates two baseline approaches: Isolation Forest as an unsupervised anomaly-detection method and XGBoost as a supervised classification method. Experiments are conducted using the reduced ROSPaCe dataset and compare two evaluation strategies: a stratified random row split and a source-file-aware holdout. The results demonstrate that evaluation strategy materially affects measured model performance. XGBoost achieved an F1-score of 0.9997 under the random row split, compared with 0.8576 under the source-file-aware holdout. The file-aware evaluation also showed a precision-recall trade-off, with XGBoost achieving 0.9985 recall and 0.7515 precision. The project is intended as an independent research and learning project for academic portfolio development and graduate research preparation. It does not claim a novel intrusion-detection algorithm. The report documents the methodology, experimental results, limitations, and future research directions. The source code and experiment files are available in the associated GitHub repository:https://github.com/Timsmall94/CPS-Intrusion-Detection