Work

GENIE-IDS: Failure-steered Closed-loop Generative-discriminative Learning for Long-tailed Network Intrusion Detection

Year: 2026
Type: article
Abstract: Rare attack families remain the weak point of learned network intrusion detection.Established remedies allocate synthetic samples by class frequency, so a class that is scarce but already separable is treated exactly like one whose examples sit inside a region occupied by a dominant class.This paper presents GENIE-IDS, in which that allocation follows the detector's own failure geometry: a calibrated model is trained first, and its margin deficit, temperature-scaled uncertainty and attraction toward competing head classes pool into a class-and-region failure map that spends a fixed budget of accepted synthetic flows on underlearned minority regions.A conditional tabular generator is queried only for those regions, and a four-stage gate screens validity, local conformity, multivariate dependence and duplication before accepted flows re-enter training with a reliability weight.Under group-aware partitions on four benchmarks, against ten baselines sharing the split, feature space, tuning allowance and both an equal accepted-sample and an equal generator-call budget, the model the declared stopping rule selects reaches macro-F1 0.789, 0.759, 0.679 and 0.868, with tail-tier gains of 4.6 to 8.0 points.Holm correction over 72 per-class blocks separates it from all ten baselines, but a dependence-aware re-analysis treating the dataset as the cluster cuts the effective sample size to about twelve, so the twenty independent dataset-seed runs are adopted as the primary inferential basis and keep the advantage over every baseline; the within-dataset Wilcoxon signed-rank test against each dataset's strongest baseline then separates the method on three of the four datasets, the exception being CSE-CIC-IDS2018 against BA-REG (p = 0.084).Removing region-level failure localisation costs more than removing any other component.
Language: English
Open Access status: hybrid
OpenAlex ID: W7222759799
FWCI:
Cites: 17
Cited by:
Related to: 10