https://jjcit.org/paper/287 10.5455/jjcit.71-1766245748 Mohamed Goismi,Mohamed Debbab,Moustafa Maaskri,Djamel Seghier Intrusion-detection system,Information gain,Grey wolf optimizer,TabNet,Deep learning,Feature selection,Hyper-parameter optimization,Network security 673 137 21-Dec.-2025 17-Feb.-2026 12-Mar.-2026 Network intrusion-detection systems (NIDSs) are critical for protecting modern cyber-infrastructure against evolving threats, yet they face persistent challenges, including high-dimensional feature spaces, class imbalance, limited interpretability and high training cost. This paper proposes IG-GWO-TabNet, a three-stage framework that (i) applies Information Gain to select a compact and discriminative feature sub-set, (ii) uses the Grey Wolf Optimizer to tune TabNet hyper-parameters over a controlled search space and (iii) leverages TabNet attention masks to provide interpretable decisions. We evaluate the approach on four public benchmarks (CIC-IDS2017, NSL-KDD, UNSW-NB15 and CIC-DDoS2019) under a leak-free protocol with stratified cross-validation, reporting both predictive performance and efficiency (training/inference cost). On CIC-IDS2017, IG-GWO-TabNet reaches 99.47 ± 0.11% accuracy and 99.46 ± 0.10% macro-F1, significantly outperforming the strongest tuned baseline (Wilcoxon signed-rank, ρ<0.001). Across datasets, the improvements remain statistically significant, while the feature-selection stage reduces runtime and supports practical deployment.