GNN-SML: Graphic Neural Network-Based Spectrum Misuser Localization

Spectrum misuser localization (SML) is essential for dynamic spectrum sharing (DSS) to ensure that only authorized users access and utilize shared spectrum. In this paper, we introduce GNN-SML, an innovative framework for crowdsourcing-based DSS systems that accurately and simultaneously localizes multiple spectrum misusers with unknown locations and transmission powers, even when they are in close proximity. GNN-SML employs location-centric, sensor-agnostic Graph Neural Networks (GNNs) with inductive power. In each online SML instance, these GNNs predict the Received Signal Strength (RSS) residue between each potential transmitter and available spectrum sensors, which may be located arbitrarily and not involved in the training phase. These predicted RSS residues, combined with real-time RSS measurements from spectrum sensors, are used with the zero-forcing technique to estimate the locations and transmission powers of all potential transmitters. This enables the identification of spectrum misusers as localized transmitters lacking proper spectrum authorizations. We validate the effectiveness and efficiency of GNN-SML with real indoor and outdoor datasets. Compared to leading methods, GNN-SML reduces localization error by up to 58% and transmitter power estimation error by up to 24%, while maintaining a comparable localization time of 1.2 s.