Data Augmentation

The foundational augmentation technique that adds controlled thermal noise to clean signal captures. By varying the Signal-to-Noise Ratio (SNR) across a defined range, the model learns to extract device-specific impairments that are robust to background interference. This simulates the varying noise floors encountered in real-world receiver deployments, preventing the network from overfitting to the pristine capture conditions of a controlled enrollment environment.

Applies synthetic multipath fading and delay spread by convolving the raw IQ samples with randomized channel models. Standard models include:

• Rayleigh fading for non-line-of-sight urban environments
• Rician fading for scenarios with a dominant direct path
• Tapped Delay Line (TDL) models for frequency-selective fading This forces the feature extractor to decouple the transmitter's hardware fingerprint from the transient channel effects, a core challenge in robust RF identification.

Introduces synthetic frequency and timing synchronization errors between the transmitter and receiver. By applying a randomized parts-per-million (ppm) offset to the carrier frequency and sampling clock, the augmentation mimics the hardware tolerances of low-cost oscillators. This is essential for ensuring that the fingerprinting model does not rely on a specific, stable frequency offset as a spurious identifier, but instead learns the deeper, invariant signal structure.

Applies non-linear temporal transformations to simulate clock jitter and Doppler shift effects. Techniques include:

• Random resampling with cubic spline interpolation
• Elastic time stretching to mimic small-scale Doppler compression
• Random cropping and zero-padding to vary burst length These augmentations teach the model to focus on the shape of transient events and steady-state modulation imperfections rather than their absolute duration or precise temporal alignment.

Artificially manipulates the complex signal constellation to generate new training samples with varying hardware impairment profiles. This includes applying a random phase rotation to simulate local oscillator mismatch and introducing controlled gain imbalance and quadrature error between the I and Q branches. By synthesizing a continuum of distortion levels, the model learns a continuous embedding space that captures the specific device's impairment signature rather than memorizing a single static constellation.

An advanced regularization strategy that creates new training samples through the convex combination of two existing IQ sequences and their labels. For RF data, this is often performed at the feature-map level rather than on raw time-domain samples to preserve physical plausibility. The technique enforces linear behavior between classes in the learned embedding space, which significantly reduces adversarial brittleness and improves generalization for open-set recognition tasks where unknown emitters must be rejected.