Hussein Fawaz, Omran Ayoub, Davide Andreoletti, Silvia Giordano
SUPSI, Switzerland
Wireless On-demand Network Systems and Services Conference (WONS), 2026
The problem
Edge IoT devices operate under strict energy constraints, yet they are increasingly expected to run reliable ML-based security mechanisms such as network intrusion detection systems (NIDS). Techniques that make these models more reliable — such as uncertainty-aware models or calibration — improve detection quality, but often increase computational cost. This raises a key question: how much energy does reliability actually cost?
What we propose
We evaluate multiple ML-based intrusion detection pipelines deployed at the edge and systematically enhance their reliability using techniques such as calibration-aware learning and uncertainty modeling. For each configuration we measure detection performance, reliability metrics, and energy consumption under realistic edge workloads.
Results and takeaways
Improving reliability consistently increases energy usage, but not always proportionally: some reliability techniques offer better energy–reliability trade-offs than others, and feature and model choices strongly influence energy efficiency. These results highlight the importance of energy-aware trustworthy AI for edge security deployments, where reliability must be balanced against limited resources. As a next step, we plan to release an open-source benchmarking framework for evaluating reliability–energy trade-offs in edge NIDS.
