Lightweight Machine Learning For Real-Time Intrusion Detection On Resource-Constrained Iot Gateways: A Comparative Study

Abstract

        We propose an energy-aware framework that uses a light-weight reinforcement learning (RL) controller to adaptively select model variants and runtime knobs (sampling rate, quantization, early-exit thresholds) for intrusion-detection tasks on resource-constrained IoT gateways. The controller balances detection utility against latency and energy cost and operates with a small profiling footprint. We evaluate the approach on standard IoT flow datasets and commodity SBCs; metrics include accuracy, F1, p50/p95 latency, model size, memory footprint and Joules per inference. The framework yields substantial energy reductions under realistic workloads while maintaining comparable detection performance to static high-accuracy models. Reproducible measurement scripts and artifacts are provided..