Domain Adaptation

Uses a gradient reversal layer (GRL) and a domain discriminator to force the feature extractor to learn domain-invariant representations. During training, the feature extractor is optimized to maximize the domain classifier's loss, effectively hiding domain-specific information.

• DANN (Domain-Adversarial Neural Network) is the foundational architecture
• Minimizes the Maximum Mean Discrepancy (MMD) between source and target feature distributions
• Particularly effective when target domain labels are entirely unavailable
• Commonly applied to adapt models trained on synthetic GNU Radio datasets to real USRP-captured signals

Measures and minimizes the statistical distance between source and target feature distributions using explicit divergence metrics. Unlike adversarial methods, this approach directly optimizes a distributional distance function.

• CORAL (Correlation Alignment) aligns second-order statistics by matching covariance matrices
• MMD-based methods minimize the distance between kernel embeddings of feature distributions
• Often combined with pseudo-labeling where high-confidence target predictions are used for self-training
• Works well when the target domain has a small set of unlabeled samples for unsupervised adaptation

Leverages auxiliary self-supervised objectives on unlabeled target data to learn domain-relevant features before or during the primary AMC task. The model learns signal structure without requiring modulation labels.

• Rotation prediction: Classify artificially rotated I/Q constellations to learn geometric signal properties
• Contrastive predictive coding: Predict future I/Q samples in a sequence to capture temporal structure
• Jigsaw puzzle solving: Reconstruct shuffled signal segments to learn local coherence patterns
• Particularly valuable when target domain SNR conditions differ significantly from source training data

Instead of adapting to a specific target domain, this technique randomizes source domain parameters during synthetic data generation to force the model to learn invariant features that generalize broadly.

• Randomizes carrier frequency offset (CFO), symbol rate, phase offset, and timing error
• Applies varying multipath fading profiles and noise levels during training
• Produces a single model robust to a wide range of deployment conditions without target-specific retraining
• Widely used in sim-to-real transfer for robotic RF systems and cognitive radio testbeds

Enables rapid adaptation to new target domains using only a handful of labeled examples per modulation class. The model learns an embedding space where samples cluster around prototypical representations of each class.

• Computes a prototype vector for each modulation class as the mean of its support set embeddings
• Classifies query samples by nearest-prototype distance in the learned metric space
• Allows field-deployed systems to quickly adapt to new hardware platforms or unseen channel conditions
• Effective when the target domain introduces new modulation schemes not present in source training

Adapts a pre-trained AMC model at inference time by updating Batch Normalization (BN) layer statistics using the incoming target data stream, without requiring any backpropagation or label information.

• Replaces source-domain BN running means and variances with target batch statistics
• Requires no gradient computation, making it suitable for real-time, low-latency deployment
• Particularly effective when domain shift is caused by hardware-specific I/Q impairments
• Can be combined with entropy minimization to further refine BN parameters during deployment