Generative AI for Wireless Interference

Jamming and other interference can cripple a wireless link, and in the field its parameters are rarely known in advance. DWSL is exploring how generative AI can learn interference directly from radio measurements and recreate it on demand, using plain-language text prompts as the controls.

Text-controlled interference synthesis

Our approach treats interference as an image-generation problem. Two software-defined radios (USRP N210s) record jamming signals over the air, either a tone or a chirp at low, medium, or high interference levels, and the receiver’s IQ samples are converted into spectrograms. We fine-tune Stable Diffusion XL, an open text-to-image model, on these spectrograms using DreamBooth and low-rank adaptation (LoRA), so that a prompt such as {chirp, low} or {tone, high} produces a new, realistic interference spectrogram.

Diagram: a transmitter radio sends a jamming signal; a receiver radio records IQ data that becomes training data; model training produces a trained model that turns a text prompt into a generated spectrogram.
How it works: measured jamming signals train a diffusion model that then generates new interference spectrograms from text prompts. Figure from Tylek et al., IEEE WAMICON 2025. © 2025 IEEE.

Unlike traditional interference models, which rely on predefined signal models, this approach needs no prior knowledge of the jammer’s parameters or the transmitted signal: it learns from labeled measurements alone.

Six spectrograms generated by the model, in two rows of three: tone interference at low, medium and high levels, then chirp interference at low, medium and high levels.
Spectrograms generated by the trained model for six prompts: (a–c) tone at low, medium, and high interference; (d–f) chirp at low, medium, and high interference. © 2025 IEEE.

Tested over the air

To check that the model captures real signal structure rather than just color patterns, we converted generated spectrograms back into IQ samples and transmitted them from a third radio, a dedicated jammer, against a live link between two radios. During each 60-second session the jammer was active for only 20 seconds. Even so, the strongest prompt, {tone, high}, raised the average packet error rate to 24.8% and cut the average throughput from 21.3 to 16.0 kB/s, while the mildest, {chirp, low}, produced a 6.8% packet error rate. The interference level named in the prompt translated into a measurable effect on the link.

Why it matters, and what’s next

Synthesizing realistic interference on demand gives us a new way to test and train resilient radios. It builds on our earlier work on mitigating jamming with pattern-reconfigurable antennas and machine learning, and on the lab’s software-defined radio infrastructure, including the Grid SDR Testbed. Next, we plan to integrate retrieval-augmented generation (RAG) so the system can adapt in real time to complex interference patterns and to signal types it hasn’t seen before.

We welcome collaborators and prospective students interested in generative AI for wireless systems. Contact Kapil Dandekar (dandekar ~ drexel.edu).

Team

Matthew Tylek, Keith Truongcao, Md Shakir Hossain, and Kapil R. Dandekar

Publication

M. Tylek, K. Truongcao, M. S. Hossain, and K. R. Dandekar, “Generative AI for Wireless Interference Modeling: Text-Controlled Waveform Synthesis Using Stable Diffusion,” in 2025 IEEE Wireless and Microwave Technology Conference (WAMICON), pp. 1–4, 2025. doi:10.1109/WAMICON64429.2025.11004115

Sponsors

This research was supported by the National Science Foundation (grant CNS-1816387) and Lockheed Martin Advanced Technology Laboratories.