A reproducible security-ML laboratory that compares anomaly classifiers across multiple traffic datasets, exposing model drift, data leakage, class imbalance, calibration errors, and cross-environment performance loss.

The benchmark standardizes ingestion, preprocessing, semantic feature mapping, feature selection, class-imbalance handling, model training, probability calibration, and experiment tracking. Logistic Regression, Decision Tree, SVM, Random Forest, and gradient-boosted models are evaluated within their source dataset and against an independent traffic distribution. Model cards report performance by attack category together with inference latency, model size, calibration quality, and explainability evidence.
