This paper explores a machine-learning approach for detecting jamming attacks in private 5G networks by framing the problem as one-class classification. Rather than relying on prior knowledge of attacker behavior, the authors aim to detect whether a received signal deviates from what is considered “legitimate,” making the method suitable for real-world deployments where jamming strategies may vary over time. The proposed solution draws on a theoretical insight: training a supervised classifier with legitimate data and an artificially generated non-legitimate class can approximate the behavior of a Generalized Likelihood Ratio Test (GLRT)—a classical statistical method for one-class detection. To implement this, the system uses a convolutional neural network (CNN) trained on (i) real 5G IQ samples representing normal operation and (ii) synthetic data uniformly distributed over the IQ domain to represent unknown anomalies.
The network setup includes a private 5G testbed with a gNodeB, user equipment, a watchdog radio capturing IQ samples, and a jammer transmitting wideband noise. Incoming IQ streams are converted into 128×128 bitmaps, providing the CNN with a spatial representation of the signal. Importantly, real jamming examples different noise and interference patterns—are used only for testing, ensuring the model learns a generalizable boundary rather than specific attack signatures. To benchmark performance, the authors compare the CNN-based GLRT approximation with a Convolutional Autoencoder (CAE) trained solely on legitimate samples. While the CAE detects anomalies through reconstruction error, the CNN explicitly learns a separation between legitimate and artificial non-legitimate data.
Experimental results show that the CNN provides clearer separation between normal and jammed samples and achieves lower false-alarm and miss-detection rates across multiple jamming types. These findings demonstrate that combining synthetic data with a supervised learning framework enables effective, model-agnostic jamming detection even when attackers use previously unseen strategies. Overall, the paper presents a practical and flexible method for enhancing the resilience of private 5G networks, showing that GLRT-inspired one-class classification can provide robust physical-layer protection without requiring prior knowledge of the adversary.
One-Class_Classification_as_GLRT_for_Jamming_Detection_in_Private_5G_Networks