An embedded testbed to measure energy and accuracy in federated anomaly detection under concept drift: special session paper

Anomaly detection at the edge increasingly relies on federated learning to share knowledge across resourceconstrained devices. At the same time, concept drift forces deployed models to adapt as the definition of normality evolves. However, such approaches are typically validated in simulation, so their energy claims are not grounded in measurements on real hardware. We present FL-STM32N6, an open-source embedded testbed in which three STM32N6570-DK microcontroller boards run the complete federated anomaly detection process on-device: inference, drift detection, and training. The firmware is baremetal Rust, and a workstation acts as FedAvg aggregation server and test harness. A source measurement unit samples the supply current of one node, and cycle counts attribute the energy cost to each phase. An 864-run campaign on three industrial datasets shows that federated retraining under drift achieves the highest accuracy in every experimental cell, with no measurable increase in device energy consumption.