Research
Class-imbalance baselines for network intrusion detection · paper under review
Role
Researcher
Duration
6 Months
Date
2025 - 2026
Type
Independent Research
Accuracy is a misleading metric when the attack you care about is 0.07% of the traffic. This study runs an 18-experiment grid: two tasks (binary and multiclass), three models (LR, RF, XGBoost), and three imbalance strategies (none, class weighting, SMOTE). It includes explicit rare-class analysis on Worms and Shellcode, reports macro-F1, ROC-AUC, and G-Mean, and fits preprocessing only on training data to prevent leakage. The resulting paper is under peer review.
Which class-imbalance strategy actually helps an IDS detect rare attacks, and does the answer change between binary and multiclass settings? Most published results hide this behind aggregate accuracy.
A fully reproducible experiment framework with fixed seeds, config snapshots, strict train/validation/test separation, and per-class metrics. The pipeline generates radar charts and faceted heatmaps directly from the results.
18
Controlled Experiments
1
Paper Under Review
100%
Reproducible (Seeded)
Reach out for source code walkthroughs or collaboration.