Cycle-Consistent GAN (CycleGAN)

Unlike conditional GANs that require matched pairs of signals, CycleGAN learns the mapping between two RF domains using unpaired datasets. One set contains simulated IQ samples, another contains real over-the-air captures—no correspondence between individual samples is needed. This is essential for sim-to-real transfer where collecting perfectly aligned signal pairs is physically impossible due to hardware and channel variations.

The defining constraint of CycleGAN is the cycle-consistency loss: a signal translated from the source domain to the target domain and back again must match the original input. For RF, this means a simulated QPSK signal translated to appear 'real' and then translated back to 'simulated' should reconstruct the original waveform. This bidirectional consistency prevents mode collapse and ensures the translation preserves the underlying modulation structure rather than producing arbitrary outputs.

CycleGAN deploys two generator networks and two discriminator networks operating in opposite directions:

• Generator G: Source → Target (e.g., simulated → real)
• Generator F: Target → Source (e.g., real → simulated)
• Discriminator D_X: Distinguishes real vs. translated source samples
• Discriminator D_Y: Distinguishes real vs. translated target samples

This symmetric architecture enables bidirectional domain adaptation, allowing models trained on synthetic data to generalize to real RF environments while preserving signal semantics.

An optional identity loss term encourages generators to preserve color and composition when inputs already resemble the target domain. For RF applications, this means a real-world signal passed through the 'simulated-to-real' generator should remain largely unchanged. This regularization prevents the generator from introducing unnecessary distortions to signals that are already in the target distribution, stabilizing training and improving fidelity for signals near the domain boundary.

In RF domain adaptation, CycleGAN learns to translate clean simulated signals into signals exhibiting realistic channel impairments without explicitly modeling the channel. The generator implicitly captures:

• Multipath fading characteristics
• Doppler spread from mobility
• Hardware impairments like IQ imbalance and phase noise
• Non-linear amplifier distortion

This learned channel model enables rapid generation of realistic training data without computationally expensive physics-based channel simulation for every sample.

The total objective combines adversarial losses from both discriminators with the cycle-consistency loss and optional identity loss. Training alternates between:

• Updating discriminators to better distinguish real from translated signals
• Updating generators to fool discriminators while minimizing cycle loss

This multi-objective optimization ensures translated RF signals are both indistinguishable from real signals (adversarial) and semantically faithful to the original modulation (cycle-consistent), preventing the generator from simply memorizing the target distribution.

The defining constraint of the CycleGAN architecture. It enforces that a signal translated from the source domain (e.g., simulated IQ) to the target domain (e.g., real-world IQ) and back again must be identical to the original input.

• Forward Cycle: x → G(x) → F(G(x)) ≈ x
• Backward Cycle: y → F(y) → G(F(y)) ≈ y This prevents mode collapse and ensures the generator does not ignore the input structure while matching the target distribution.

The broader machine learning discipline that CycleGAN serves. It addresses the distribution shift between a labeled source domain (e.g., synthetic MATLAB signals) and an unlabeled target domain (e.g., over-the-air captures).

• CycleGAN performs unsupervised domain adaptation by learning a mapping without paired examples.
• Alternative techniques include Gradient Reversal Layers (GRL) for domain-invariant feature extraction.
• Critical for closing the sim-to-real gap in RF machine learning deployments.

An alternative to CycleGAN when paired data is available or when specific class control is required. A cGAN conditions both the generator and discriminator on auxiliary information, such as modulation type or signal-to-noise ratio (SNR).

• Unlike CycleGAN, cGANs typically require aligned input-output pairs.
• Useful for generating specific classes of synthetic RF signals on demand.
• Often combined with adaptive discriminator augmentation (ADA) to prevent overfitting on small RF datasets.

A sim-to-real transfer strategy that does not rely on adversarial translation. Instead, the parameters of the simulated RF environment—such as noise floor, delay spread, and Doppler shift—are randomized widely during training.

• Forces the model to learn features invariant to channel conditions.
• Often simpler to implement than CycleGAN but requires a highly parameterized channel simulator.
• Complements CycleGAN when used to pre-train a model before fine-tuning with cycle-consistent adaptation.

An emerging class of generative models that synthesize high-fidelity RF waveforms by iteratively denoising random Gaussian noise. They represent a powerful alternative to GAN-based translation.

• Do not suffer from mode collapse, a common GAN failure where the generator produces limited variety.
• Can be conditioned for domain translation tasks similar to CycleGAN.
• Currently more computationally intensive at inference time compared to single-pass generator networks.