Overview
A university term paper executed with research-grade discipline: an 18-experiment grid crossing two tasks (binary and multiclass) with three models (LR, RF, XGBoost) and three strategies (no balancing, class weighting, SMOTE) on the UNSW-NB15 dataset. Preprocessing fits only on training data to prevent leakage, seeds and config snapshots are fixed, and the results archive includes Friedman nonparametric tests, metric confidence intervals, per-class metrics, and rare-class aggregates for Worms (0.07%) and Shellcode (0.65%), moving well past raw accuracy with macro-F1, ROC-AUC, and G-Mean. A one-click Colab notebook reruns the study, and IEEE-style paper drafts with a professor review report live in the private repo.