Benchmarking Unsupervised Jamming Detection in 5G Open RAN
| Source: arXiv AI
Tags: Open RAN, 5G, Jamming Detection, Autoencoders, LSTM, VAE
A benchmark of unsupervised jamming detectors on 5G Open RAN telemetry across ten environments finds a statistical ensemble method beats autoencoders, with mean F1 of 0.918 versus 0.521.
Details
RF jamming threatens 5G networks, and Open RAN telemetry allows detection without dedicated spectrum monitoring. Labeled attack data is rare, which pushes operators toward unsupervised detectors that are hard to compare. The authors introduce a detector-independent composite score combining detection confidence and temporal interval coherence, and use it to compare a Statistical Ensemble Method against Vanilla, LSTM and VAE autoencoder baselines. On KPI telemetry from ten heterogeneous 5G environments with fixed and variable jamming bursts, SEM averaged a composite of 0.906 under leave-one-out evaluation versus 0.688 for the best autoencoder. Ground-truth checks gave SEM a mean F1 of 0.918 versus 0.521. The composite ranked SEM first in every environment, suggesting simpler statistical detectors are a strong baseline for RAN security monitoring.