This paper proposes a lightweight wireless intrusion detection method that identifies 5G jamming attacks using spectrograms derived directly from received IQ samples. Instead of relying on higher-layer metrics like PER or SINR which can fluctuate naturally and often fail to reveal targeted jamming the approach analyzes power spectral density (PSD) patterns at the physical layer, enabling reliable detection even at very low SINR levels. To evaluate the method, the authors build an experimental 5G standalone system (n78 band, 100 MHz bandwidth) and generate both legitimate and jammed signals using software-defined radios. Spectrograms are constructed by stacking PSDs computed from 1024-sample FFT windows, forming 100×1024 matrices that capture fine-grained frequency–time structure.
Two learning models are compared:
-
Unsupervised CAE, trained only on non-jammed spectrograms, detects anomalies through reconstruction error.
-
Supervised CNN, trained on both jammed and non-jammed data, performs direct classification.
Both achieve high accuracy, with the CAE showing clean separation between jammed and normal cases (reconstruction error ≈50× higher for jammed samples), and the CNN providing the lowest FA/MD rates. Execution time is low in both cases, under 48 ms for CAE and under 46 ms for CNN in 95% of trials, enabling fast response to attacks. The results demonstrate that spectrogram-based learning models can detect broadband jammers accurately and efficiently, without requiring changes to 5G standards or interaction with the network making them suitable for standalone watchdog devices.
Detecting_5G_Signal_Jammers_Using_Spectrograms_with_Supervised_and_Unsupervised_Learning