ARISE: A daptive R esource-Aware I ntelligent SE nsing Enabling Intermittent Learning

The application of Federated Learning (FL) to battery-free, energy-harvesting devices is fundamentally challenged by intermittent power and the high cost of data labeling. Standard FL protocols are impractical for such devices, as they assume continuous operation and fully labeled datasets. We proposeARISE, an intelligent sensing scheme that integrates Adaptive Resource-aware sampling with a reactive intermittent computing model to overcome these limitations.ARISEco-trains and federates a task-aware active learning sampler, enabling each client to intelligently select and annotate only the most informative local data samples. This process is managed by a lightweight, reactive state machine that ensures computational progress across frequent power failures. Evaluations on CIFAR-10, CIFAR-100, and Tiny-ImageNet show thatARISEachieves superior accuracy with up to 40% fewer labeled samples while reducing communication overhead by over 25%.ARISEprovides a robust and efficient solution for deploying privacy-preserving AI on the growing ecosystem of battery-free edge devices.