Spec-GNN: Spectrum Enforcement Through Graph Neural Networks in Dynamic Spectrum Access Systems

The underlay access mode in dynamic spectrum access (DSA) systems permits secondary users to transmit concurrently with the primary user, provided that the cumulative interference imposed on the primary user does not exceed a set threshold. A spectrum outage is said to have occurred when the interference threshold has been exceeded. To limit the occurrence of an outage, the spectrum sharing policy mandates that all secondary users transmit within set power limits. However, an outage could still occur due to “spectrum violators” who are secondary users that fail to adhere to the spectrum policy, or it could occur due to unforeseen noise within the spectrum environment. In this work, we design an algorithm for detecting and identifying spectrum violators (if any), which we collectively term "enforcement". We propose a novel graph neural network (GNN) based algorithm, Spec-GNN, to identify which secondary users, if any, are spectrum violators when an outage occurs. Because of the noise in a communication system, outages can occur even without the presence of violators, and thus a key challenge is to keep the false alarm rate low. Spec-GNN performs by utilizing as input, a graph of the DSA system formed from data collected from monitoring sensors deployed in the environment. Spec-GNN then learns the roles of each secondary user in the graph, allowing it to classify them as violators or not. We extensively evaluate Spec-GNN across diverse settings with varying number of available sensors, secondary users, and violators. The results show that even with low sensor densities, Spec-GNN can achieve accuracy of around 95% with false alarm rates as low as under 0.03 in realistic outage scenarios. We also show that Spec-GNN’s performance is quite robust to the amount of participating secondary users in the DSA system. Even when the number of violators makes up as much as 50% of the secondary users, Spec-GNN is still able to achieve a classification accuracy of close to 92%, while keeping the false alarm rate under 0.04.