FedSensor: Federated Learning Framework for Secure Sensor-Based IoT at the Extreme Edge

Ultra-low-power IoT devices at the extreme edge generate valuable data for machine learning applications, however, these devices face stringent limitations in processing power, memory, and energy capacity. The current paper introduces FedSensor, a federated learning framework specifically designed to securely deliver global model updates and enable inference on highly constrained devices, without requiring firmware alterations. The architecture safeguards both data and metadata through end-to-end encryption and enforces edge-based device management to enhance system resilience. FedSensor attains energy efficiency by delegating training tasks to edge servers while maintaining localized inference at the device level. Experimental results demonstrate that battery longevity is preserved, with energy consumption increasing by less than 8% for inference intervals of 10 seconds, and remaining below 5% for longer intervals. These results substantiate the viability of deploying secure, autonomous, and energy-efficient federated learning on ultra-low-power IoT devices.