Answers are grounded in this paper and cite their sources.
Fixed single-model deployment can be restrictive for edge intrusion detection when traffic conditions and available computational resources change after deployment. This study proposes RC-PORT-IDS, a two-stage model-portfolio planning framework with observable-state recourse. The first stage selects a constrained portfolio of heterogeneous intrusion-detection models before the operating state is observed. The second stage selects a feasible deployed model using observable traffic and resource states only. Seven tabular classifiers were evaluated. Canadian Institute for Cybersecurity Internet of Things 2023 dataset was used for development analysis, including duplicate-conflict control, observable-regime construction, resource benchmarking, and internal recourse-gain analysis. Canadian Institute for Cybersecurity Internet of Medical Things 2024 dataset was used for independent external replication with retraining, three-fold out-of-fold regime-complete calibration, policy freezing, and final test evaluation. Internal development produced positive recourse gain in 13 of 15 deployment cases, with a mean of 0.001512. In the external replication, the calibration-selected constrained static policy obtained a mean balanced accuracy of 0.697275, a constrained test-oracle static policy obtained 0.840868, and full traffic-plus-resource recourse obtained 0.888692. Full recourse therefore retained a mean advantage of 0.047825 over the test-oracle static policy. Traffic-only recourse reached 0.837476, while resource-only recourse remained at 0.697275. Sensitivity analyses over regime count, regime seed, and ±5% latency-threshold perturbations preserved positive recourse gain in 12 of 15 cases. The results show that a resource-constrained pre-deployed portfolio can support state-dependent model selection under controlled storage and latency constraints without using ground-truth attack-family information at runtime.
No takes yet. Share an insight, caveat, or question.
Shaheen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: