Domain Classifier

The domain classifier is trained to maximize its ability to distinguish between source and target domains, while the feature extractor is simultaneously trained to minimize the classifier's accuracy. This creates a minimax game where the feature extractor learns to strip away domain-specific information, producing representations that are indistinguishable across domains. The gradient reversal layer is a common implementation technique.

Typically implemented as a shallow multi-layer perceptron (MLP) attached to the bottleneck of a feature extractor via a gradient reversal layer. The classifier receives the feature vector and outputs a softmax distribution over domain labels. Key design choices include:

• Depth: Usually 2-3 fully connected layers
• Width: Often 256-1024 hidden units
• Placement: Attached to the deepest, most abstract feature layer

The domain classifier is trained with standard cross-entropy loss against the true domain labels. However, during backpropagation to the feature extractor, the gradient is reversed and scaled by a hyperparameter λ. This forces the feature extractor to maximize domain classification error. The total loss is: L_total = L_task - λ * L_domain, where λ controls the trade-off between task performance and domain invariance.

A pseudo-function that acts as an identity transform during forward propagation but multiplies the gradient by -λ during backpropagation. This eliminates the need for alternating optimization steps. The GRL is inserted between the feature extractor and the domain classifier, ensuring that features become domain-agnostic while the classifier itself continues to learn domain discriminability.

In RF fingerprinting, the domain classifier is trained to predict the channel condition (e.g., indoor vs. outdoor, specific room, multipath profile) under which a signal was captured. By adversarially suppressing this information, the feature extractor learns to focus on hardware-intrinsic impairments rather than propagation artifacts. This enables a model trained in one environment to authenticate devices in entirely different channel conditions.

The λ hyperparameter controls the strength of domain adversarial pressure. Common scheduling strategies include:

• Fixed λ: Simple but requires careful tuning
• Progressive increase: λ starts at 0 and grows according to 2 / (1 + exp(-γ * p)) - 1, where p is training progress
• Adaptive λ: Adjusted based on the ratio of task loss to domain loss to maintain balanced training

The process of separating learned representations into independent, interpretable factors of variation. In RF fingerprinting, the goal is to isolate:

• Device-specific factors: DAC non-linearities, I/Q imbalance, oscillator phase noise
• Channel-specific factors: Multipath delay spread, Doppler shift, path loss

A Domain Classifier enforces this separation by ensuring the primary feature space contains no information useful for predicting the channel domain, effectively purging channel artifacts from the device signature.