Mixup

Mixup generates a new training sample (x̃, ỹ) by taking a convex combination of two random samples from the training batch. The mixing coefficient λ is sampled from a Beta(α, α) distribution, typically with α ∈ [0.1, 0.4]. For RF data, this operates directly on raw IQ samples: x̃ = λ * x_i + (1 - λ) * x_j. The label is mixed identically: ỹ = λ * y_i + (1 - λ) * y_j. This forces the model to transition linearly between classes rather than making abrupt, brittle decisions.

By training on interpolated signal examples, Mixup penalizes models that oscillate sharply between RF classes. The technique minimizes the vicinal risk rather than the standard empirical risk, effectively imposing a regularization constraint that the model's output should change linearly between any two training points. This is particularly valuable for automatic modulation classification where signal constellations exist on a continuous manifold and adversarial perturbations exploit sharp decision boundaries.

Unlike standard augmentation that preserves hard one-hot labels, Mixup produces soft targets that represent the proportional membership of the synthetic sample to each source class. For a two-class mix with λ=0.7, the label becomes [0.7, 0.3] rather than [1, 0]. This implicit label smoothing reduces model overconfidence and has been shown to improve calibration in RF fingerprinting tasks where emitter identities exist on a spectrum of hardware similarity.

Mixup-trained models demonstrate significantly improved resilience against adversarial attacks on RF systems. By learning to behave linearly in the vicinity of training samples, the model's gradient is smoothed, reducing the effectiveness of gradient-based evasion attacks such as the Fast Gradient Sign Method (FGSM). In electronic warfare contexts, this means a Mixup-trained signal classifier is less likely to be fooled by carefully crafted jamming waveforms designed to cross decision boundaries.

Mixup is applied on-the-fly during mini-batch construction in the training loop. For complex-valued IQ data, the interpolation is performed independently on the in-phase (I) and quadrature (Q) components. A critical implementation detail for RF is ensuring that the mixed signal's power is normalized post-interpolation to prevent the creation of unrealistic amplitude levels. Frameworks like PyTorch and TensorFlow support Mixup natively through custom collate functions or dedicated augmentation layers.

When combined with channel impairment simulation, Mixup helps models generalize across diverse propagation conditions. By interpolating between signals subjected to different Rayleigh fading profiles or Doppler shifts, the model learns a continuous representation of the channel space. This reduces the sim-to-real gap by preventing the model from memorizing specific channel realizations present in the training set, a common failure mode in over-parameterized neural receivers.

A transfer learning paradigm that bridges the simulation-to-reality gap by aligning feature distributions between a labeled source domain (e.g., synthetic RF) and an unlabeled target domain (e.g., live over-the-air captures).

• Uses Gradient Reversal Layers (GRLs) to learn domain-invariant representations
• Critical when real-world labeled RF data is unavailable
• Often combined with Mixup to smooth decision boundaries across domains

A dual-network architecture where a generator synthesizes realistic RF waveforms while a discriminator attempts to distinguish them from real signals. This adversarial game produces high-fidelity synthetic training data.

• WGAN variants use Wasserstein distance to prevent mode collapse
• Conditional GANs generate specific modulation types on demand
• Provides unlimited labeled data for Mixup interpolation

A sim-to-real strategy that deliberately varies simulation parameters—noise floor, delay spread, Doppler shift—across extreme ranges during training. This forces the neural network to discard domain-specific features and latch onto invariant signal characteristics.

• Prevents overfitting to specific channel conditions
• Pairs effectively with Mixup for continuous interpolation across randomized environments

A self-supervised pre-training method that learns robust RF representations without labels. The model pulls augmented views of the same signal together in embedding space while pushing apart representations of different signals.

• Uses techniques like SimCLR and MoCo adapted for IQ samples
• Creates a structured latent space where Mixup interpolations remain semantically meaningful
• Enables strong performance in few-shot RF classification scenarios

A regularization technique that injects maliciously perturbed RF examples into the training set to harden models against evasion attacks. Unlike Mixup's convex combinations, adversarial training exposes the model to worst-case perturbations near the decision boundary.

• Defends against adversarial interference in contested spectrum environments
• Often used jointly with Mixup for complementary robustness guarantees

A semi-supervised technique where a model generates high-confidence predictions on unlabeled RF captures and treats them as ground truth for iterative retraining. This expands the effective training set without manual annotation.

• Applies confidence thresholds to filter noisy pseudo-labels
• Mixup can be applied between labeled and pseudo-labeled samples to smooth the semi-supervised decision boundary
• Essential for leveraging massive volumes of unlabeled spectrum captures