Adversarial robustness does not equal zero-day generalization in intrusion detectors
Original title: Adversarial Robustness Is Not Zero-Day Generalization: A Controlled Evaluation of Flow-Based Intrusion Detectors on Corrected CIC-IDS2017
Perturbation-based robustness tests overestimate zero-day detection; models maintain 90.9-99.7% under attacks but drop to 22.8-76.2% on held-out families.
The paper tests whether synthetic variants of known attacks proxy for zero-day attacks in NIDS evaluation. On corrected CIC-IDS2017, Random Forest, XGBoost, 1D-CNN and FT-Transformer show high detection under constrained perturbations yet low detection on leave-one-attack-family-out tests. Robustness exceeds zero-day generalization by 24-72 percentage points. Similar gaps recur on CSE-CIC-IDS2018.
- Robustness 90.9-99.7% vs LOAO 22.8-76.2%
- Gaps of 24-72 percentage points
- Tree ensembles confirmed by evasion search
- Results hold on temporal splits
From the original article
1 min read at arxiv.orgSynthetic variants of known attacks, including feature-space adversarial perturbations, are sometimes used as stand-ins for unseen ("zero-day") attacks when evaluating machine-learning network intrusion detection systems (NIDS). We test whether this proxy holds.
By Abdulhamid Zaro · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.