A Knowledge-Driven Graph Convolutional Network for Spectrum Prediction

In the 6G mobile communication systems, the integration of sensing, communication, and array signal processing technologies heralds a new era of advanced situational awareness. However, this advancement raises critical concerns about privacy and security, especially with the widespread deployment of devices equipped with radar-like sensing capability, including malicious ones. Moreover, suffering from rapidly changing spectrum environments and the restricted capability of spectrum prediction, jamming effectiveness evaluation values may not be fully obtained. Therefore, in order to effectively jam the malicious sensing device and improve the spectrum utilization, spectrum prediction is expected to be a key enabler, which features predicting the jamming effectiveness evaluation values and recommending frequency-domain subbands for utilization. In order to characterize the relation between jamming devices and malicious sensing devices when predicting spectrum, we construct a spectrum knowledge graph (SKG). In order to encode the KGE vectors from the triplets composed of entities and relations, we develop the CNN-based KGE model, where the CNN can be exploited to derive embedding encoded with knowledge of relations. In order to address the problem of predicting spectrum, we convert this issue into a node regression problem based graph model and graph convolution network (GCN). In order to address the problem of sparse feature vectors related to entities and relationships within the SKG, we exploit KGE vectors generated by a CNN-based KGE algorithm to replace the one-hot encoding method, which can be used to formulate the feature matrix in the graph model. Simulation results indicate that the spectrum prediction algorithm based on a knowledge-driven graph convolutional network (KGCN) can significantly improve the accuracy of spectrum prediction.