Conditional GAN (cGAN)

The core architectural innovation of a cGAN is the injection of a conditioning variable y (e.g., modulation type label, SNR value, or device ID) into both the generator and discriminator. This is typically achieved by concatenating the condition with the input noise vector z in the generator and with the input signal in the discriminator. This forces the generator to learn the conditional distribution P(x|y) rather than the marginal distribution P(x), enabling on-demand synthesis of specific signal classes without the need for post-generation filtering or sorting.

cGANs solve the fundamental problem of uncontrolled generation in standard GANs. By conditioning on a one-hot encoded modulation label, a cGAN can be directed to produce a clean QPSK waveform, then immediately switch to generating a 16-QAM signal with identical channel impairments. This capability is critical for building balanced RF datasets where rare modulation schemes are underrepresented. The generator learns to disentangle the class-specific signal structure from the stochastic noise, producing high-fidelity, labeled examples on demand for training downstream automatic modulation classification (AMC) models.

Beyond discrete class labels, cGANs can be conditioned on continuous physical parameters such as signal-to-noise ratio (SNR), carrier frequency offset, or Doppler shift. By feeding a scalar SNR value as the condition, a single trained generator can synthesize signals across a continuous range of noise levels. This allows for the creation of datasets with fine-grained SNR granularity, which is essential for training robust AMC models that must operate across varying link budgets. The generator learns a smooth manifold where interpolating between SNR values produces physically plausible intermediate signal distortions.

In a cGAN, the discriminator receives both the signal and its condition, learning to evaluate not just realism but conditional consistency. It must determine whether a QPSK signal truly exhibits the phase transitions of QPSK or if it is a mismatched 16-QAM sample. This dual objective prevents the generator from ignoring the condition and producing a single plausible class. The discriminator's loss function includes an auxiliary classification component, effectively turning it into a multi-task critic that enforces both fidelity and semantic correctness of the generated RF waveform.

cGANs can be conditioned on channel metadata to generate signals with specific propagation effects. By conditioning on a power delay profile identifier or a Rician K-factor, the generator learns to apply realistic multipath fading to the baseband waveform. This enables the creation of paired datasets where the same transmitted signal is observed under multiple known channel conditions. Such datasets are invaluable for training channel estimation neural networks and for evaluating the robustness of physical layer algorithms to specific, controlled impairments without requiring exhaustive over-the-air data collection campaigns.

A prominent cGAN variant for RF applications is the Auxiliary Classifier GAN (AC-GAN). In this architecture, the discriminator outputs both a realism score and a class label prediction. The generator is optimized to maximize the probability of the correct class, providing a stronger training signal than simple concatenation. For RF fingerprinting, an AC-GAN can be conditioned on device identity, learning to generate signals that carry the unique hardware impairments of a specific transmitter. This enables the augmentation of emitter-specific datasets for training specific emitter identification (SEI) systems.

The foundational architecture upon which cGANs are built. A GAN pits a generator against a discriminator in a minimax game. The generator learns to map random noise to realistic data, while the discriminator learns to distinguish real samples from fakes. In RF applications, this adversarial process produces synthetic IQ samples that mimic real transmissions. The cGAN extends this by feeding conditioning information to both networks.

A transfer learning paradigm that addresses the sim-to-real gap. When a cGAN is trained on simulated RF data but deployed on over-the-air captures, the statistical mismatch degrades performance. Domain adaptation techniques align the feature distributions of the source and target domains. Key methods include:

• Adversarial domain adaptation using a gradient reversal layer
• Maximum Mean Discrepancy (MMD) minimization
• Correlation alignment of second-order statistics

A GAN variant that replaces the binary cross-entropy loss with the Wasserstein distance (Earth Mover's Distance). This provides a smoother, more meaningful gradient signal to the generator, dramatically improving training stability. For RF signal generation, WGAN mitigates mode collapse, ensuring the synthetic dataset captures the full diversity of modulation schemes, SNR levels, and channel impairments present in the real distribution.

An unpaired image-to-image translation architecture adapted for RF domain transfer. CycleGAN learns to translate signals from a simulated domain to a realistic domain without requiring matched pairs of examples. It enforces cycle-consistency loss: a signal translated from source to target and back must remain identical. This is critical when you have abundant simulated data but only unlabeled real-world captures.

A catastrophic failure mode in GAN training where the generator learns to produce only a narrow subset of the target distribution. In RF synthesis, a collapsed cGAN might generate only QPSK signals regardless of the conditioning label requesting 64-QAM. Mitigation strategies include:

• Minibatch discrimination to encourage diversity
• Wasserstein loss for stable gradients
• Adaptive Discriminator Augmentation (ADA) to prevent discriminator overfitting on small datasets

A self-supervised pre-training strategy that learns robust representations without labels. The model pulls positive pairs (augmented views of the same signal) together in embedding space while pushing negative pairs (different signals) apart. When combined with a cGAN, contrastive pre-training on unlabeled RF data produces a conditioning-aware latent space that dramatically improves generation quality with limited labeled examples.