Machine Learning for Non-Intrusive Profiling of IoT Devices in Noisy Environments

This paper investigates the impact of noise on the performance of a side-channel-based system for classifying operational scenarios in Internet of Things (IoT) devices. The approach discriminates eight usage scenarios executed on a Raspberry Pi by exploiting physical information leaked during operation, supporting non-intrusive behavioral monitoring and constituting a preliminary step toward side-channel-based intrusion detection in realistic deployment conditions.The acquisition architecture relies on two complementary sensing channels: an electromagnetic probe for radiated emissions and a current probe for power consumption. From the acquired signals, representative time- and frequency-domain features are extracted and used to train and evaluate a machine learning-based classifier.Starting from a baseline framework validated under controlled laboratory conditions, this work evaluates how measurement noise affects classification performance. Noise is injected at the raw signal level and controlled via the signal-to-noise ratio (SNR), considering multiple degradation levels to emulate practical acquisition disturbances typically encountered in real environments. Accuracy is assessed for electromagnetic-only, current-only, and combined configurations to compare channel robustness and feature-fusion effectiveness.Results show a progressive degradation of classification accuracy as noise increases, with electromagnetic features more sensitive than current-based measurements. However, combining both sensing modalities consistently improves robustness across all tested conditions. These findings quantify minimum signal-quality requirements and highlight robustness constraints for reliable side-channel-based scenario classification in realistic IoT environments.