CycleGAN RF Augmentation

Unlike supervised pix2pix models, CycleGAN learns to translate between source and target RF domains without requiring paired examples. This is critical for RF augmentation because capturing perfectly aligned simulated and over-the-air signals is often impractical. The model learns the mapping through cycle consistency: translating a signal from domain A to B and back to A should recover the original signal.

• Eliminates the need for expensive paired data collection
• Enables translation between simulated IQ and real over-the-air captures
• Learns bidirectional mappings simultaneously (A→B and B→A)

The core innovation of CycleGAN is the cycle consistency constraint, which enforces that translating a signal from source to target domain and back should reconstruct the original. For RF signals, this means a simulated IQ sample translated to appear as over-the-air, then translated back to simulated, must match the original.

• Forward cycle: x → G(x) → F(G(x)) ≈ x
• Backward cycle: y → F(y) → G(F(y)) ≈ y
• Prevents mode collapse and ensures meaningful domain transformation
• Preserves signal structure while altering domain-specific characteristics

Two PatchGAN discriminators operate on each domain, classifying whether signal representations are real or generated. The generator learns to produce RF signals indistinguishable from real captures, while the discriminator learns to detect synthetic artifacts.

• Discriminator A distinguishes real over-the-air signals from translated simulations
• Discriminator B distinguishes real simulated signals from translated over-the-air captures
• Adversarial training drives realistic channel impairment modeling
• Forces generators to learn multipath fading, noise, and hardware imperfections

An optional identity mapping loss regularizes the generator to preserve signal content when the input already belongs to the target domain. When a real over-the-air signal is fed to the generator that simulates over-the-air effects, it should remain unchanged.

• Prevents unwanted modulation distortion during translation
• Preserves spectral occupancy and bandwidth characteristics
• Maintains temporal envelope of the original signal
• Critical for preserving modulation scheme identity across domain shifts

CycleGAN excels at transferring complex channel impairments between domains without explicit channel modeling. The generator implicitly learns to apply realistic multipath profiles, Doppler shifts, phase noise, and non-linear distortions characteristic of the target domain.

• Transforms clean simulated IQ into faded, noisy over-the-air captures
• Transfers hardware-specific impairments between different receiver front-ends
• Enables augmentation across different SNR regimes
• Learns compound impairment distributions without analytical models

The generator typically employs a U-Net or ResNet-based encoder-decoder architecture adapted for complex-valued IQ signals. Downsampling layers capture spectral context while skip connections preserve fine-grained temporal structure essential for modulation fidelity.

• Instance normalization adapts to per-sample channel statistics
• Residual blocks enable deep architectures without degradation
• Complex-valued convolutions preserve phase relationships in IQ data
• Skip connections maintain transient signal features across translation

A semi-supervised learning principle that enforces a model to produce similar predictions for an unlabeled data point and its perturbed or augmented versions. This concept is central to CycleGAN's cycle-consistency loss, which ensures that translating a signal to another domain and back yields the original input.

• Cycle-consistency loss: ||G(F(x)) - x||
• Purpose: Prevents mode collapse and preserves signal structure
• RF application: Ensures modulation type and bit content survive domain translation

The task of identifying RF signal inputs that differ fundamentally from the training data distribution. When CycleGAN augments training data with synthetic samples, robust OOD detection becomes critical to ensure the model recognizes when it encounters truly novel emitters or unknown modulation schemes.

• Open-world spectrum monitoring: Must flag unknown signals
• Augmentation risk: Synthetic data may not cover all real-world edge cases
• Techniques: Energy-based models, Mahalanobis distance in feature space

A data augmentation strategy that creates virtual training samples by linearly interpolating raw IQ sequences and their corresponding labels. This promotes linear behavior between training examples and improves generalization, often used alongside CycleGAN-generated data to further regularize RF classifiers.

• Formula: x̃ = λx_i + (1-λ)x_j
• Label mixing: ỹ = λy_i + (1-λ)y_j
• Benefit: Encourages smoother decision boundaries in signal space