A routing-signal study of confidence-gated conditional computation for lightweight IoT intrusion detection

Abstract

Background Intrusion detection at Internet of Things (IoT) edge gateways must run under tight compute and memory budgets, motivating dual-path designs that classify most flows with a cheap model and escalate only uncertain flows to a heavier one. A prior conference study introduced such a system, in which a tiny multilayer perceptron (MLP) on features compressed by principal component analysis (PCA) is the fast path, gated to a heavy MLP by the maximum softmax probability (MSP). That system recovers near-heavy accuracy while escalating few flows. Methods We extend that work into a systematic study of what signal should drive the escalation gate. On three benchmarks (ACI-IoT-2023, TabularIoTAttacks-2024, and the large-scale CICIoT2023), we compare routing signals (MSP, predictive entropy, margin, energy, temperature-scaled confidence, latent norm) against a learned router and an oracle bound at matched cost; study temperature-scaling calibration and a leakage-free threshold rule; test a tiny-ensemble disagreement signal; and evaluate robustness, latency, model size, statistical significance, and cross-dataset transfer. Results Confidence-gated routing gives large, statistically significant gains over the tiny model on every dataset, whereas a learned router improves only marginally over plain MSP, so once an escalation budget is fixed a simple MSP gate suffices. On CICIoT2023 we observe an oracle gap. Routing cannot reach the heavy model even at large budgets, because the tiny model is confidently wrong on a few high-volume attack families such as denial-of-service (DoS) floods that are never escalated. Neither ensembling nor confidence-independent signals close the gap. A fixed threshold does not transfer across datasets, its escalation rate swinging from 1 % to 55 % . Conclusions The efficiency benefit grows with a stronger heavy backbone, including a gradient-boosted-tree heavy. Gates should be set by a budget, and closing the oracle gap requires identifying confident errors within high-confidence classes rather than a better routing signal.