Answers are grounded in this paper and cite their sources.
The rapid expansion of the Internet of Things (IoT) has widened the network attack surface, leaving infrastructure vulnerable to high-dimensional threats.While deep learning (DL) is an effective defense, existing hybrid models struggle with class imbalance, feature-scaling distortions, and poor minority-class handling under standard loss functions.AEGIS-Net is a CNN-BiLSTM-MHA classifier intended as an explainable and generalizable multi-class intrusion detection system; its generalization is tested rather than assumed.Feature preprocessing consists of quantilebased non-parametric feature scaling, selective SMOTE oversampling and categorical focal loss.Performance is evaluated against CICIoT2023 dataset (34 classes) and cross-dataset generalization to CICIoT-DIAD2024 and CICIoMT2024 datasets.Explainable models use Integrated Gradients, KernelSHAP, GradientSHAP and LIME.For five different random seeds, AEGIS-Net obtains 99.08% ±0.02% accuracy, macro-F1 = 0.7481 ± 0.0067 and minority F1 = 0.4845 ± 0.0128, while the cross-entropy/StandardScaler baseline provides 97.93%, 0.6473 and 0.2966 correspondingly.The improvements stem from quantile-based scaling (F1 gain = 0.053) and SMOTE (F1 gain = 0.027).Under the same split, preprocessing, five seeds and early stopping, random forest (macro-F1 0.8207 ± 0.0054) and a Tab-Transformer (0.7685 ± 0.0093) outperform AEGIS-Net (0.7481 ± 0.0067), whereas AEGIS-Net is comparable to XGBoost (0.7418 ± 0.0087) and exceeds LightGBM (0.7305 ± 0.0072).The LIME surrogate fidelity is low (median R² = 0.27; only 11% have R² ≥ 0.80).The top five feature overlaps between four different attribution methods are low (they have 1.4-2.3common features).In zero-shot transfer, the CICIoMT2024 score is 63.84% and CICIoT-DIAD2024 is 15.61%.When trained on labeled 10% dataset, they have scores of 86.60% and 64.67%, respectively, and XGBoost scores 99.78% and 75.69%.Source pre-training offers no significant improvement in accuracy using 50 samples per class.When the source model is retrained on the directly observable common features only, zero-shot accuracy is 70.36% (macro-F1 0.1445) on CICIoMT2024 and 9.11% (macro-F1 0.0473) on CICIoT-DIAD2024, so the transfer failure is not explained by missing features alone.
No takes yet. Share an insight, caveat, or question.
Soni et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: