Work

Multi-Type Flow-Level Intrusion Detection Using 2Z-RSTOA Feature Selection and a BCGM-WPNN Hybrid Ensemble: A Benchmark Evaluation on NSL-KDD and UNSW-NB15 Towards MANET Deployment

Year: 2026
Type: article
Abstract: Mobile ad hoc networks (MANETs) have no fixed infrastructure, so intrusion detection must run on the nodes themselves. This paper proposes a flow-level, multi-type intrusion detection module intended for a trusted centralized MANET node and evaluates it on two wired-network benchmarks; MANET operation itself is not evaluated. Flow features are filtered with a chi-square test and reduced by a hybrid Zebra–Zaslavsky–Radial Siberian Tiger optimizer (2Z-RSTOA). Minority classes are balanced with SMOTE on training data only, an adaptive neuro-fuzzy module (ANSPIS) produces rule scores, and a bagged ensemble of DNN, DBN and MLP learners, combined by Gaussian mean weighting and a Phish-activated CNN meta-learner (BCGM-WPNN), assigns the attack family. In stratified 10-fold cross-validation, BCGM-WPNN reached 79.49 ± 3.71% macro-F1 on five-class NSL-KDD and 54.88 ± 1.66% on ten-class UNSW-NB15. This was the highest among the neural models but significantly below a Random Forest on both datasets. On the official test sets, macro-F1 was 56.53 ± 3.67% on KDDTest+, on par with the best baseline, and 45.04 ± 1.00% on UNSW-NB15. On the UNSW-NB15 testing set the model detected 99.49% of attack flows but labelled 39.50% of normal flows as attacks, a false-positive rate too high for unattended deployment. The measurable benefit of the ensemble lies in rare-class and denial-of-service detection; in ablation, the Phish activation, Gaussian weighting and SMOTE gave no measurable gain. Code, settings and per-run results are released for replication.
Source: Array
Language: English
Open Access status: gold
Estimated APC paid: $1,680
OpenAlex ID: W7226019645
FWCI:
Cites: 33
Cited by:
Related to: 10