Work

Empirical Evaluation of a Dual-Pillar DDoS Detection Framework: Temporal Causality, Feature Representations, and Operational Trade-Offs

Year: 2026
Type: preprint
Abstract: Operational Distributed Denial of Service (DDoS) defense requires timely volumetric monitoring and low-false-alarm flow classification, yet offline evaluations often obscure how temporal ordering, feature representation, and class balance govern real-world performance. This paper evaluates a dual-pillar DDoS framework on the two-day CICDDoS2019 benchmark (50.06M Training-Day and 20.36M Testing-Day flows). Pillar 1 examines an ordinary differential equation (ODE)-inspired residual monitor on populated-row packet- and byte-rate series. Contrasting label-concatenated central differencing with chronological backward differencing shows that apparent early warnings stem from benign bursts and forward look-ahead, while mean-shifted absolute scoring inflates ROC-AUC without lowering active false-positive rates (FPR); on the chronological trace, raw-volume z-scoring and resetting CUSUM yield markedly lower FPR (1.37–3.82%) than the residual transfer (12.19–22.43%). Pillar 2 benchmarks 12 classifiers, three normalizers, and seven feature selectors against a combined 2D Haar pipeline across four Testing-Day views. Combined 2D Haar + Extra Trees achieves the highest attack recall (99.970–99.996%, 99.242–99.255% accuracy), whereas raw Min-Max SVM (99.680%, 0.508% FPR) and 12-feature Lasso Logistic Regression (99.652%, 0.554% FPR) achieve higher accuracy and cut false positives by 65.8% and 62.7%, respectively, with Lasso more than doubling prevalence-adjusted precision at 0.1% attack prevalence.
Language: English
Open Access status: green
OpenAlex ID: W7236714356
Cites:
Cited by:
Related to: 10