Radio frequency signal generation using denoising diffusion implicit models
Addison Nute, Matthew Judah, Scott Kuzdeba · 2025
The capture, processing, and storage requirements of radio frequency (RF) datasets impose a difficult challenge when developing, testing, and validating algorithms in dynamically evolving RF-environments. Neural Network based diffusion models have achieved remarkably realistic results in the domain of image and video generation. However, these models require a large computing overhead of processing noise for many steps to produce a high-quality example and have been optimized using techniques not optimized for the RF domain. In this work, we demonstrate that denoising diffusion implicit modes with noise schedules optimized for IQ data types can be used to produce realistic communication signals in the RF domain without the necessary computing used in denoising diffusion probabilistic models (DDPMs). The approach utilizes an UNET-based autoencoder and DDIM implementation to produce quality RF sample generation, with meaningful features for future RF algorithm development.