Generative Adversarial Network (GAN)

The core mechanism is a two-player zero-sum game defined by a minimax loss function. The generator (G) minimizes the objective while the discriminator (D) maximizes it, formalized as min_G max_D V(D, G). The discriminator is trained to maximize the probability of assigning the correct label to both real training examples and samples from G, while G is simultaneously trained to minimize log(1 - D(G(z))), effectively maximizing the probability of D making a mistake.

The generator is a differentiable function, typically a deep deconvolutional neural network, that maps a latent vector z sampled from a simple prior distribution (e.g., Gaussian noise) to the data space. In the RF domain, this output is a synthetic IQ sample vector. The generator learns to model the true data distribution p_data(x) without ever seeing it directly, instead receiving gradients through the discriminator's classification error.

The discriminator is a binary classifier, often a convolutional neural network, that outputs a scalar probability estimating the likelihood that a given sample came from the real training data rather than from the generator. It acts as a learned, adaptive loss function. As the generator improves, the discriminator must learn increasingly subtle features to distinguish real from synthetic RF waveforms, driving both networks toward higher fidelity.

Training ideally converges when the generator perfectly replicates the true data distribution and the discriminator is forced to random guessing, outputting 1/2 for all inputs. At this Nash equilibrium, the Jensen-Shannon divergence between the real and generated distributions is minimized. In practice, this equilibrium is notoriously difficult to reach due to the non-convex, non-stationary nature of the optimization landscape.

A critical failure mode where the generator discovers a narrow set of outputs that reliably fool the discriminator and ceases to explore the full data distribution. In RF synthesis, this manifests as a GAN producing only a single modulation variant instead of the full constellation. Mitigation strategies include Wasserstein loss, unrolled GANs, and mini-batch discrimination to enforce output diversity.

In a Conditional GAN (cGAN), both the generator and discriminator are augmented with auxiliary information y, such as a modulation type label or target signal-to-noise ratio. This transforms the model from an uncontrolled generative process into a steerable one. The objective becomes a conditional two-player minimax game, enabling the targeted synthesis of specific RF signal classes on demand for balanced dataset creation.

A GAN variant that replaces the standard Jensen-Shannon divergence with the Wasserstein distance (Earth Mover's Distance) to provide a smoother, more meaningful loss metric. This stabilizes training by eliminating vanishing gradients and provides a loss curve that correlates with sample quality. For RF applications, WGANs are critical for generating complex, multi-modal signal distributions without suffering from mode collapse, ensuring the synthetic dataset captures the full diversity of real-world waveforms.

An architecture that conditions both the generator and discriminator on auxiliary information, such as modulation type, signal-to-noise ratio (SNR), or a specific emitter ID. This allows for controlled synthesis of specific RF signal classes on demand. In RF data augmentation, a cGAN can be directed to produce thousands of unique QPSK samples at a target SNR of 10 dB, directly addressing class imbalance in a training set without generating unwanted signal types.

An unpaired translation architecture that learns to map signals between two domains without requiring matched pairs of data. It enforces cycle-consistency loss: a signal translated from a simulated domain to a real domain and back must remain identical. This is uniquely powerful for closing the sim-to-real gap in RF. A CycleGAN can learn to translate clean, simulated waveforms into realistic over-the-air captures by adding subtle hardware impairments and channel artifacts, all without needing perfectly aligned real-world examples.

A catastrophic training failure where the generator learns to produce only a single, or very few, varieties of output that successfully fool the discriminator, ignoring the full diversity of the target distribution. In an RF context, a collapsed GAN might generate only one specific pulse shape repeatedly, regardless of the intended modulation. Mitigation strategies include Wasserstein loss, unrolled GANs, and adaptive discriminator augmentation (ADA) to force the generator to explore the entire signal manifold.

A dual-purpose technique where a model is hardened by injecting maliciously perturbed examples into the training set. In the GAN context, this is the core learning dynamic. In the broader RFML context, adversarial training defends against evasion attacks on the physical layer. A classifier trained with adversarial RF examples—signals with imperceptible, crafted noise designed to cause misclassification—becomes robust to both intentional jamming and unintentional channel distortions that mimic adversarial perturbations.

A training stabilization technique that dynamically applies a range of augmentations to both real and generated samples before they enter the discriminator. By adaptively controlling the augmentation probability based on overfitting heuristics, ADA prevents the discriminator from memorizing a small real RF dataset. This is essential for RF data augmentation where real-world captures are scarce; ADA enables stable GAN training on as few as a few thousand real signal examples without the discriminator becoming overconfident.