Deep Learning for Anomaly Detection and Classification in Radio-Frequency Spectra

Sammanfattning

Reliable communication, navigation, and surveillance are fundamental to modern critical infrastructure, yet radio-frequency (RF) environments are increasingly exposed to unintentional interference and deliberate attacks such as jamming and spoofing. This thesis develops an open-world deep-learning framework for RF anomaly monitoring from spectrogram images, combining modules for unsupervised anomaly detection and localisation, known-anomaly classification, open-set unknown rejection, and clustering-based absorption of new anomaly classes. The full open-world classification and absorption protocol is evaluated on Synthetic Anomaly Signals (SAS), while the authentic IAD benchmark is used as an external anomaly-detection test. On the mid-SNR SAS setting, PatchCore reaches 99.94% image-level AUROC and the one-round open-world pipeline achieves 94.48% holdout accuracy. PatchCore also outperforms the Conv-VAE baseline on IAD detection. Overall, the results suggest that deep visual models are promising for RF anomaly detection, classification, and class discovery, while the synthetic-only scope of the open-world results motivates further evaluation on more realistic RF streams before autonomous deployment.