FedLoss: In-Region Location Verification in 6G Networks via Personalized Federated Learning

In wireless networks, in-region location verification (IRLV) refers to the problem of verifying whether a transmitting device is inside a region of interest, based on the channel estimated from the received signal. In this paper, we consider multiple regions of interest and multiple base stations (BSs), each locally performing IRLV with a machine-learning (ML) model on estimated channels. We exploit the spatial correlation of channels and adopt a federated learning strategy among the BSs. Still, we also want to obtain different local models to reflect the different statistics of observed channels. We propose FedLoss, a personalized federated learning framework tailored to physical-layer IRLV under non-IID channel conditions. FedLoss operates in two phases: a federated averaging stage that learns a shared representation across BSs, and a locally regularized fine-tuning stage that adapts the global model to each BS's channel statistics. A loss-based switching criterion determines the transition between the two phases, enabling efficient and stable training. We evaluate the proposed approach using realistic 3GPP-inspired channel models in a multi-base-station 6G scenario.