Convolutional and Variational Autoencoders with Supervised Classifiers for RF Spectrum Anomaly Detection

Effective and secure monitoring of the radio frequency spectrum is essential for the reliable operation of modern wireless communication systems. Automated detection of spectral anomalies can support regulatory and operational monitoring tasks. This study evaluates three autoencoder architectures — Vanilla Autoencoder (AE), Convolutional Autoencoder (CAE), and Variational Autoencoder (VAE) — for anomaly detection using waterfall image representations derived from FM-band (86.5–108 MHz) measurement data. The models were assessed both as standalone reconstruction-based detectors and as feature extractors combined with supervised classifiers. Standalone autoencoders achieved F1-scores between 0.794 and 0.813. Higher performance was obtained when encoder-derived latent representations were used with supervised models, where the best result (F1-score: 0.915) was achieved by the CAE + ExtraTrees combination. The results indicate that hybrid encoder–classifier approaches can provide an effective practical solution for anomaly detection in spectrum monitoring environments. While promising, the findings are limited to the investigated FM-band dataset and require further validation across broader spectrum environments.