Diffusion Models

Unlike GANs that generate samples in a single forward pass, diffusion models synthesize RF waveforms through a Markov chain of successive refinement steps. Starting from pure Gaussian noise, the model iteratively removes noise over hundreds or thousands of timesteps, gradually revealing a coherent signal structure. This process allows the model to correct errors at each step, resulting in high-fidelity IQ samples that faithfully capture the statistical nuances of real-world transmissions. The iterative nature provides fine-grained control over the generation trajectory.

Diffusion models circumvent the adversarial training instability that plagues GANs, such as mode collapse and non-convergence. The training objective is a simple regression task: predicting the noise added to a clean signal at a given timestep. This provides a stable, well-behaved loss landscape. For RF applications, this stability is critical when modeling complex signal distributions with multiple modulation types and varying signal-to-noise ratios, eliminating the need for careful balancing of generator and discriminator networks.

The synthesis process can be guided by conditioning information injected at each denoising step. A conditional diffusion model can generate specific RF waveforms based on auxiliary inputs such as:

• Modulation type (QPSK, 16-QAM, OFDM)
• Signal-to-noise ratio (SNR) target
• Channel impairment profiles (delay spread, Doppler shift) This enables the creation of labeled, diverse synthetic datasets tailored to specific operational scenarios, directly addressing class imbalance and rare-event simulation for signal intelligence training.

By operating in a compressed latent space—as seen in Latent Diffusion Models—the computational cost of generating high-dimensional RF waveforms is dramatically reduced. The model learns a perceptually equivalent but lower-dimensional representation of IQ samples. Sampling from this latent space allows for stochastic interpolation between signal classes and the generation of novel, yet physically plausible, waveforms that lie off the manifold of the original training data, enhancing model robustness against unknown signal variants.

Diffusion models excel at learning complex data distributions, making them powerful tools for domain translation. A diffusion model can be trained to translate simulated RF waveforms into realistic over-the-air equivalents by learning the residual distortion characteristics of physical hardware and propagation environments. This process effectively adds the unmodeled physical imperfections—such as IQ imbalance, phase noise, and non-linear amplifier effects—that constitute the sim-to-real gap, without requiring paired examples of simulated and real signals.

This technique balances sample diversity and fidelity without requiring a separate classifier model. During training, the conditioning signal (e.g., modulation label) is randomly dropped, teaching the model to generate both conditionally and unconditionally. At inference, the output is pushed toward the conditioned prediction and away from the unconditional one using a guidance scale. For RF data augmentation, this produces synthetic signals that are both highly representative of the target class and sufficiently diverse to prevent overfitting in downstream classifiers.