Data Augmentation

The most fundamental augmentation for RF signals. Varying levels of AWGN are added to clean IQ samples to simulate different Signal-to-Noise Ratio (SNR) conditions.

• Trains the classifier to be robust to thermal noise and channel degradation.
• Prevents the model from overfitting to high-SNR, lab-collected signals.
• Example: Augmenting a 30dB SNR sample with noise to create synthetic 10dB and 5dB versions.

Applying synthetic fading, frequency offset, and phase rotation to clean baseband signals mimics real-world propagation effects.

• Rayleigh/Rician Fading: Simulates multipath reflections in urban environments.
• Carrier Frequency Offset (CFO): Introduces a slight mismatch between transmitter and receiver oscillators.
• Phase Noise: Models the instability of local oscillators.
• This forces the feature extractor to learn invariances to non-ideal channel conditions.

Label-preserving transformations that exploit the translation equivariance properties of convolutional neural networks.

• Time Shifting: Cyclically rolling the IQ sequence forward or backward in time. This simulates random packet arrival times and burst start positions.
• Frequency Shifting: Applying a small, controlled frequency translation to the complex baseband signal. This simulates residual down-conversion errors.
• Both techniques dramatically increase the effective dataset size without altering the underlying modulation scheme.

Advanced interpolation-based techniques that create new synthetic samples by blending existing ones.

• Manifold Mixup: Performs linear interpolation on the learned hidden representations of signals, not the raw IQ samples. This creates smoother decision boundaries in the embedding space.
• Input Mixup: Linearly combines two raw IQ samples and their corresponding labels: x_new = λ*x_a + (1-λ)*x_b.
• These methods act as a strong regularizer, encouraging the classifier to behave linearly between training examples and reducing adversarial vulnerability.

Using a Generative Adversarial Network (GAN) trained on real signals to synthesize entirely new, plausible IQ samples for underrepresented modulation classes.

• The generator learns the underlying probability distribution of the signal constellation and can sample novel variations.
• Particularly useful for rare modulation types where only a handful of real captures exist.
• Can be combined with a conditional setup to generate samples at specific target SNR levels or channel conditions.

Instead of augmenting raw IQ data, this technique synthesizes new feature vectors in the learned embedding space of a pre-trained encoder.

• A generator network is trained to produce plausible feature representations for novel classes based on a small support set.
• Distribution Calibration: The statistics (mean, covariance) of base classes are used to calibrate the estimated distribution of a novel class, from which new features are sampled.
• This is computationally cheaper than raw signal generation and directly targets the classification boundary.

Creates realistic, artificially-generated RF waveforms using generative models like GANs or VAEs to supplement limited real-world training data for rare modulation types. Unlike basic transformations, this approach learns the underlying data distribution and samples entirely new, plausible signal instances.

• GANs pit a generator against a discriminator to produce high-fidelity IQ samples
• VAEs learn a latent representation enabling controlled generation of signal variants
• Critical for simulating rare modulation schemes with no over-the-air captures

A data augmentation strategy operating in the feature space rather than the raw signal domain. A generator network synthesizes additional, plausible feature vectors for underrepresented classes to improve the class boundary in few-shot learning.

• Augments learned embeddings instead of raw IQ samples
• Preserves semantic structure while expanding class support
• Often combined with prototypical networks for refined class prototypes

A regularization and augmentation method that performs linear interpolations not on raw inputs but on learned hidden representations. This encourages smoother decision boundaries and better generalization in few-shot modulation classifiers.

• Interpolates between hidden states of different signal classes
• Produces continuous blends that lie on the data manifold
• Reduces overfitting when training data is extremely scarce

A statistical technique that calibrates the feature distribution of base classes to estimate the distribution of novel classes. This enables the generation of high-quality synthetic samples for few-shot tasks without requiring any real examples from the target class.

• Transfers distributional statistics from data-rich to data-poor classes
• Assumes similar variance structure across related modulation types
• Enables effective augmentation even in zero-shot scenarios

A self-supervised representation learning method that extracts useful features from high-dimensional data by training an autoregressive model to predict future latent representations using a probabilistic contrastive loss. The learned representations serve as a robust foundation for downstream augmentation.

• Learns representations that capture long-range signal structure
• Contrastive objective separates signal from noise in latent space
• Pre-trained CPC encoders improve few-shot classifier initialization

Preprocessing and model-based techniques for mitigating fading, noise, and offset effects before classification. While not augmentation in the traditional sense, robust compensation expands the effective diversity of training data by ensuring the classifier focuses on modulation-specific features rather than channel artifacts.

• Includes carrier frequency offset correction and equalization
• Augments model robustness to real-world propagation conditions
• Essential bridge between synthetic training data and over-the-air deployment