MixUp IQ

MixUp generates a new training sample by taking a linear interpolation between two randomly drawn IQ sequences and their one-hot encoded labels. For a pair of IQ samples (x_i, y_i) and (x_j, y_j), the virtual sample is created as x̃ = λx_i + (1-λ)x_j and ỹ = λy_i + (1-λ)y_j, where λ ~ Beta(α, α). The hyperparameter α controls interpolation strength:

• α → 0: approaches standard empirical risk minimization (no mixing)
• α = 1.0: uniform sampling from the Beta distribution
• α > 1: favors mixing near 0.5, creating more ambiguous examples This operates directly on complex-valued baseband signals, preserving the mathematical structure of I/Q components during augmentation.

Unlike standard empirical risk minimization (ERM) which trains on discrete data points, MixUp implements vicinal risk minimization (VRM) by constructing a distribution of virtual examples in the vicinity of the training data. This encourages the classifier to transition linearly between class decision boundaries rather than learning sharp, brittle transitions. Key benefits for RF applications:

• Reduces memorization of spurious correlations in training data
• Smooths the decision landscape in high-dimensional IQ space
• Provides a form of implicit regularization without explicit weight penalties
• Particularly effective when training data is limited or class distributions are imbalanced

MixUp has been empirically shown to improve out-of-distribution generalization and adversarial robustness in RF modulation classification tasks. By training on interpolated samples, the model learns to produce softer, more calibrated confidence scores rather than overconfident predictions on ambiguous signals. Documented improvements include:

• Reduced calibration error (ECE) on corrupted or noisy IQ samples
• Increased robustness to channel impairments not seen during training
• Better performance on few-shot modulation recognition when combined with self-supervised pre-training
• Mitigation of label noise in real-world signal collections where annotations may be imperfect

While both are mixing-based augmentations, they operate on fundamentally different principles:

MixUp IQ:

• Blends entire IQ sequences via global pixel-wise interpolation
• Creates soft, blended labels proportional to λ
• Preserves global signal structure but creates synthetic intermediate signals

CutMix IQ:

• Replaces a rectangular region of one IQ sample with a patch from another
• Labels mixed proportionally to the area ratio of the cut region
• Forces the model to attend to less discriminative local features

MixUp is generally preferred for preserving the temporal coherence of modulated signals, while CutMix excels at improving attention to subtle hardware impairments in RF fingerprinting tasks.

MixUp is frequently applied as a strong augmentation within self-supervised learning frameworks for RF. In contrastive methods like SimCLR and MoCo, MixUp can be used alongside standard augmentations (noise addition, frequency shift, time cropping) to create diverse positive pairs. In BYOL and Barlow Twins, mixing between different unlabeled IQ samples encourages the encoder to learn representations that are invariant to linear interpolation between signal classes. This is particularly valuable for:

• Pre-training on large unlabeled spectrum captures
• Learning continuous manifolds of modulation types
• Enabling smooth interpolation between known and novel emitter classes

When applying MixUp to raw IQ data, several domain-specific factors must be addressed:

• Complex-valued mixing: Interpolation must respect the complex nature of IQ samples; λ is applied identically to both I and Q components
• Power normalization: Mixed samples may have altered signal power; re-normalization to a consistent level prevents the model from exploiting amplitude artifacts
• Phase continuity: Direct sample-wise mixing can create phase discontinuities at the boundaries of mixed signals, which may introduce unrealistic artifacts
• Batch-level mixing: MixUp is typically applied per-batch by shuffling indices and mixing pairs, adding negligible computational overhead
• Alpha scheduling: Some practitioners linearly decrease α during training to transition from strong regularization to fine-grained discrimination