Poster: HeteroRF: Heterogeneity-Aware Federated Radio-Frequency Human Activity Recognition

Ubiquitous radio-frequency (RF) devices constantly gather environmental and human data, generating substantial information for Human Activity Recognition (HAR). This data remains underutilized due to its distributed nature, privacy concerns, and lack of labels. Federated learning (FL) offers a promising solution for RF-HAR data mining, but conventional FL methods often fail due to divergence from inherent heterogeneity across distributed RF devices. We propose HeteroRF, a novel FL framework addressing RF-HAR divergence at three levels: a data augmentation mechanism at the raw-data level, curriculum learning to stabilize local training, and Gumbel softmax-based client selection to address global heterogeneity. Experiments demonstrate that HeteroRF significantly improves RF-HAR accuracy.