Synthetic Signal Generation

A dual-network architecture where a generator creates synthetic IQ samples from random noise, while a discriminator attempts to distinguish them from real signals. Through adversarial training, the generator learns to produce waveforms statistically indistinguishable from authentic captures.

• Key variants: Conditional GANs (cGANs) allow generation of specific modulation types by conditioning on a class label
• RF-specific challenge: GANs must capture both temporal structure and spectral constraints unique to modulated signals
• Training stability: Wasserstein GANs with gradient penalty (WGAN-GP) are preferred for RF data to avoid mode collapse

A probabilistic generative model that learns a compressed latent space representation of signal distributions. The encoder maps IQ samples to a Gaussian latent distribution, while the decoder reconstructs waveforms from sampled latent vectors.

• Continuous latent space: Enables smooth interpolation between modulation types and controlled signal morphing
• Disentanglement: β-VAE variants encourage separation of modulation parameters (symbol rate, carrier offset) into independent latent dimensions
• Uncertainty quantification: VAEs naturally model the probability density of signals, providing confidence bounds on generated samples

A class of generative models that learn to reverse a gradual noising process. Starting from pure Gaussian noise, the model iteratively denoises toward a coherent signal structure, producing high-fidelity waveforms.

• Denoising Diffusion Probabilistic Models (DDPMs): The foundational framework using Markov chains of noise addition and learned removal
• Score-based variants: Learn the gradient of the log probability density (score function) for more flexible sampling
• RF advantages: Diffusion models excel at capturing fine-grained spectral details and avoid GAN training instability

Generative architectures that model signals as sequential data, predicting each IQ sample conditioned on all previous samples. These models decompose the joint probability distribution into a product of conditional distributions.

• WaveNet adaptations: Dilated causal convolutions originally designed for audio synthesis have been adapted for complex-valued RF waveforms
• Transformer-based: Self-attention mechanisms capture long-range dependencies across symbol periods
• Sample quality: Autoregressive models produce highly coherent signals but suffer from slow sequential sampling at inference time

A generative approach that transforms a simple base distribution (e.g., Gaussian) into a complex signal distribution through a sequence of invertible transformations. Unlike GANs or VAEs, flows provide exact likelihood computation.

• Exact density estimation: Enables precise evaluation of how well generated samples match the training distribution
• Real NVP and Glow: Popular flow architectures using affine coupling layers adapted for complex-valued data
• Limitation: Require architectures where dimensionality is preserved throughout, constraining design flexibility for high-dimensional IQ streams

Hybrid architectures that incorporate domain knowledge of signal physics directly into the generative process. Rather than learning purely from data, these models embed constraints from communications theory.

• Spectral mask enforcement: Loss functions penalize generated signals that violate regulatory emission boundaries
• Constellation constraints: Embed known symbol geometries (QPSK, 16-QAM) as structural priors in the latent space
• Channel simulation integration: Combine generative models with differentiable channel models to produce signals with realistic impairments (fading, Doppler, phase noise)

An adversarial framework where a generator network creates synthetic IQ samples and a discriminator network attempts to distinguish them from real captures. Through iterative competition, the generator learns to produce highly realistic waveforms that capture the subtle statistical properties of target modulation schemes.

• Architecture: Typically uses deep convolutional layers tailored for complex-valued IQ data
• Training instability: Mode collapse and non-convergence are common challenges requiring careful hyperparameter tuning
• Applications: Generating rare modulation types like 256-QAM or custom military waveforms for classifier training

A probabilistic generative model that encodes input signals into a latent distribution and decodes samples from that distribution back into the signal space. Unlike GANs, VAEs provide a principled likelihood framework and a smooth, continuous latent space ideal for interpolation between known modulation types.

• Reparameterization trick: Enables backpropagation through the stochastic sampling process
• KL divergence loss: Regularizes the latent space toward a standard Gaussian prior
• Advantage: More stable training than GANs, with explicit control over latent representations

A technique where the generative model receives a class label or attribute vector as an additional input, enabling controlled synthesis of specific modulation types on demand. This is essential for targeted augmentation of underrepresented classes in a few-shot support set.

• Conditional GANs (cGANs): Feed the label into both generator and discriminator
• Conditional VAEs (CVAEs): Concatenate the condition to the encoder input and decoder latent code
• Use case: Generating exactly K additional samples for a rare modulation class with specific SNR and channel impairment parameters

Realistic synthetic signals must incorporate physics-based channel effects to bridge the sim-to-real gap. This includes multipath fading, additive white Gaussian noise (AWGN), carrier frequency offset, phase noise, and Doppler shift.

• Rayleigh and Rician fading: Simulate urban and line-of-sight propagation environments
• Hardware impairments: Model power amplifier nonlinearity and I/Q imbalance for authentic RF fingerprinting
• Importance: Without accurate channel modeling, classifiers trained on synthetic data fail catastrophically on over-the-air captures

While both techniques expand training datasets, they operate on fundamentally different principles. Data augmentation applies label-preserving transformations (rotation, noise injection, time warping) to existing samples. Generative synthesis creates entirely new, statistically plausible samples from a learned distribution.

• Augmentation: Computationally cheap but limited to transformations of existing data
• Synthesis: Can produce novel combinations of signal characteristics not present in the original dataset
• Synergy: Both techniques are often combined—augmenting synthetic outputs further diversifies the training set

A statistical technique that calibrates the feature distribution of base classes to estimate the distribution of novel classes, enabling the generation of high-quality synthetic samples for few-shot tasks. This method transfers distributional statistics from data-rich classes to data-scarce ones.

• Process: Compute mean and covariance statistics of base class features, then calibrate using statistics from the few available novel class samples
• Benefit: Produces more accurate synthetic features than generic generative models when only K=1 or K=5 real samples exist
• Application: Generating synthetic feature vectors for prototypical network training