Side-Channel Measurements and Machine Learning for Classifying Application-Level Scenarios in IoT Contexts

This work investigates the feasibility of using a general-purpose measurement system and effective classification techniques for identifying application-level scenarios on a Raspberry Pi 4 platform through electromagnetic and current sidechannel attacks. A wide measurement campaign was conducted over eight scenarios, including web browsers, email clients, and protocol-based services (HTTP, HTTPS, FTP, FTPS). The measurements were processed using RMS-based downsampling, optionally normalised with a Robust Scaler, and analysed with and without feature extraction. A Random Forest classifier, combined with Recursive Feature Elimination (RFE), was employed to enhance accuracy and reduce dimensionality. Results show that combining current and magnetic measurements with statistical feature engineering and selection achieves high classification accuracy-reaching up to $\mathbf{9 5. 8 3 \%}$-highlighting the feasibility of behavioural profiling through side-channel analysis in cyber security contexts.