Efficient Generative Wireless Anomaly Detection for Next Generation Networks

Anomaly detection in wireless signals through multi-sensor fusion has numerous real-world applications including spectrum monitoring and awareness, fault detection, and spectrum security. As networks, multi-user access schemes, and spectral density increase beyond 5G and into 6G, especially in difficult shared-spectrum and unlicensed-spectrum bands, monitoring of activity and anomalies on the air interface is a critical enabler for optimizing spectrum access, ensuring the quality of service, and automating orchestration. In this paper, we describe the problem of high-level spectrum anomaly monitoring using metadata derived from high-rate radio signals in a scalable, unsupervised, and bandwidth-friendly system, and we introduce several baselines and generative methods for interpreting this metadata into a high-level view of the air interface environment. We utilize three different anomaly detection methods, each making use of the advantages of different state-of-the-art deep learning techniques, in order to detect a set of anomalous activities in these metadata feeds caused by underlying activities in several radio bands. We evaluate performance by looking at the receiver operating characteristics of the anomaly detectors, and each of the three methods produces an AUROC and AUPRC score of >0.8 on average on different anomaly datasets.