VULNERABILITY OF DEEP NETWORK INTRUSION DETECTORS TO EVASION VARIES WITH THE TRAINING SEED
- Department of Computer Engineering, Mamou Institute of Technology, Mamou, P.O. Box 63, Republic of Guinea.
0 Downloads
6 Views
Abstract
Evaluations of deep network intrusion detectors usually rest on one trained model, a random train–test split, and an attacker that learns nothing from one flow to the next. We drop these three habits on the corrected CIC-IDS-2017. A chronological split lowers the macro F1 of a 1D CNN by only 0.042 but raises its false-positive rate from 0.003% to about 2%, mostly on one benign service seen on the last day. Nine CNNs that differ only in their seed reach a test macro F1 of 0.87 to 0.95, yet a decision-based attacker that may only pad, delay, or add packets to its own flows gets through 0% to 99% of a family’s flows depending on the detector, which explains 96% to 100% of the sum of squares across 15 attack conditions. Thirty-three MLPs behave alike, and a budget five to ten times smaller keeps the ranking of the detectors. Detection metrics do not predict this vulnerability, whereas attacking validation flows does (r_S = 0.60 to 1.00). Vulnerability should be measured over several training runs and used to choose the deployed model.
Index Terms: Intrusion detection, deep learning, adversarial evasion, decision-based attack, temporal evaluation, underspecification.
Keywords
How to Cite This Article
Ibrahima Toure et,al (2026); VULNERABILITY OF DEEP NETWORK INTRUSION DETECTORS TO EVASION VARIES WITH THE TRAINING SEED, International Journal of Advanced Research (IJAR), 14 (11), 535-548, ISSN 2320-5407.
Corresponding Author
Article Analytics
Similar Articles
3
373
0
50
3
426
0
339
1
312
This work is licensed under a Creative Commons Attribution 4.0 International License.
