Feature Embedding

Effective embeddings maximize the distance between distinct transmitter classes in the latent space. This is typically enforced through angular margin losses such as ArcFace or CosFace, which impose a multiplicative angular penalty on the logits to push class boundaries apart. The result is a hyperspherical embedding where each known device occupies a narrow, well-defined cone.

• Enables clear rejection thresholds for unknown emitters
• Reduces confusion between similar hardware models
• Critical for open space risk minimization

All samples from the same physical transmitter must map to a tight, cohesive cluster regardless of channel conditions, modulation schemes, or transmit power levels. Contrastive learning frameworks achieve this by explicitly pulling positive pairs (same device, different captures) together while pushing negative pairs apart.

• Compensates for multipath fading and Doppler shift
• Ensures consistent device identity across diverse operational scenarios
• Measured by intra-class covariance trace in the embedding space

The embedding must discard channel-specific artifacts while preserving hardware-specific impairments. Domain adversarial training pits a feature extractor against a channel classifier, forcing the network to learn representations that are invariant to propagation effects. This ensures the fingerprint, not the environment, drives the embedding.

• Uses gradient reversal layers to strip channel signatures
• Essential for mobile or vehicular emitter scenarios
• Validated through cross-environment enrollment and test protocols

Every embedding vector should be accompanied by a reliable uncertainty estimate. Evidential deep learning places a Dirichlet distribution over class probabilities, outputting both a prediction and an evidence-based uncertainty score. High uncertainty signals either a noisy capture or a genuinely unknown emitter.

• Distinguishes epistemic uncertainty (unknown device) from aleatoric uncertainty (poor SNR)
• Enables graceful rejection rather than forced misclassification
• Supports conformal prediction for guaranteed coverage rates

The embedding geometry must be aligned with a meaningful distance function for open set scoring. Mahalanobis distance, which accounts for class-conditional covariance, outperforms Euclidean distance when clusters are anisotropic. The distance to the nearest class centroid or prototype serves as the primary open set rejection statistic.

• Mahalanobis distance normalizes by cluster shape and orientation
• Prototypical networks use cosine distance to class means
• Extreme Value Theory calibrates distance thresholds for unknown rejection

Hardware impairments drift slowly over time due to temperature variation and component aging. The embedding space must accommodate this temporal evolution without triggering false unknown rejections. Continual learning with elastic weight consolidation or episodic memory replay allows the model to adapt to gradual drift while retaining knowledge of previously enrolled devices.

• Tracks slow-moving class centroids via exponential moving averages
• Prevents catastrophic forgetting during incremental enrollment
• Maintains separation between drifting known classes and true unknowns

In a well-trained embedding space, signals from the same physical transmitter naturally cluster together despite variations in modulation payload, channel conditions, or ambient noise. This occurs because the neural network learns to amplify hardware impairment signatures—such as I/Q imbalance and oscillator phase noise—while suppressing irrelevant content variations.

• Cosine similarity between embedding vectors serves as a direct authentication score.
• Clustering algorithms like DBSCAN or HDBSCAN can autonomously discover the number of active emitters in a monitored spectrum without prior enrollment.
• Embedding drift over time can be tracked to detect device aging or cloning attempts.

Feature embeddings enable open set emitter recognition by providing a natural geometry for rejection logic. Known transmitters occupy compact, bounded regions in the embedding manifold. An unknown emitter's embedding vector will fall far from any known class centroid, triggering a rejection.

• Mahalanobis distance from class prototypes accounts for intra-class covariance, providing a more reliable rejection score than Euclidean distance.
• Extreme Value Theory (EVT) can be applied to the distribution of distances within known classes to calibrate statistically rigorous rejection thresholds.
• This approach eliminates the need to train on adversarial or unknown examples beforehand, a critical requirement for dynamic spectrum environments.

Embedding networks trained with metric learning objectives—such as triplet loss or prototypical networks—learn a universal distance function for RF signals. This allows new transmitter classes to be enrolled with as few as 1-5 example transmissions.

• The network extracts a prototypical embedding for the new device by averaging the embeddings of its few enrollment captures.
• Subsequent authentications compare new captures to this stored prototype using a simple distance check, requiring no model retraining.
• This capability is essential for rapid IoT onboarding and tactical coalition operations where pre-collecting large datasets for every potential emitter is infeasible.

A primary challenge in RF fingerprinting is that multipath propagation and channel fading can distort the signal more than the hardware impairments themselves. Contrastive learning frameworks address this by explicitly training the embedding space to be invariant to channel effects.

• Positive pairs are created by passing the same transmitted signal through diverse simulated channel models; the network is penalized if their embeddings diverge.
• Domain adversarial training pits a channel discriminator against the embedding network, forcing the extraction of channel-invariant features.
• The resulting embeddings represent only the transmitter-intrinsic hardware signature, enabling reliable cross-environment authentication.

Embedding spaces transform spectrum monitoring from a frequency-and-power paradigm into a semantic search problem. An operator can query a database of historical emissions using an example signal, and the system retrieves all past transmissions from the same device or device type based on embedding proximity.

• Vector databases optimized for approximate nearest neighbor (ANN) search enable millisecond-latency queries across billions of stored embeddings.
• This enables retrospective analysis: "Show me every transmission this specific rogue emitter has made in the past month."
• Embedding-based indexing also enables zero-shot emitter type classification by measuring similarity to known device model prototypes.

High-dimensional embedding vectors can be projected into 2D or 3D for human analysis using techniques like t-SNE or UMAP. These visualizations provide intuitive, actionable intelligence for spectrum analysts.

• Distinct emitter clusters are immediately visible, revealing the number and relative similarity of active devices.
• Anomalous outliers—potentially representing spoofing attempts or malfunctioning hardware—appear as isolated points far from any known cluster.
• Embedding visualizations serve as a powerful explainability tool for building operator trust in automated RF decision-making systems, bridging the gap between neural network inference and human situational awareness.

The process of learning a distance function that assigns small distances to semantically similar pairs and large distances to dissimilar pairs. In open set emitter recognition, this is critical for rejection logic:

• Mahalanobis Distance: Accounts for feature covariance, measuring distance in standard deviations from a class mean
• Cosine Similarity: Measures angular separation, robust to magnitude variations in signal strength
• Euclidean Distance: Simple geometric distance in the embedding space

Learned metrics ensure that unknown emitters naturally fall far from all known class centroids, enabling reliable novelty detection.

A family of loss functions that enforce discriminative constraints on the angular space of feature embeddings to maximize inter-class separation. Key variants:

• ArcFace: Adds an additive angular margin penalty between features and their class weights
• CosFace: Applies a cosine margin directly to the logits
• SphereFace: Multiplies the angle by a margin factor

These losses produce embeddings where known emitter classes form tight, well-separated clusters with large angular gaps. Unknown devices naturally fall into these gaps, simplifying open set rejection thresholds.

A self-supervised representation learning framework that pulls semantically similar samples together and pushes dissimilar samples apart in the embedding space. Core mechanisms:

• Positive Pairs: Different augmented views of the same signal or transmissions from the same device
• Negative Pairs: Transmissions from different emitters or noise samples
• InfoNCE Loss: Maximizes mutual information between positive pairs while minimizing it for negatives

Contrastive pretraining produces embeddings that capture subtle hardware impairment signatures without requiring labeled device identities, ideal for few-shot enrollment scenarios.

A few-shot learning architecture that classifies query samples by computing distances to a prototypical representation of each class in a learned metric space. The prototype is simply the mean embedding of the support set for each class:

• Query embeddings are compared to all prototypes via Euclidean distance
• A SoftMax over negative distances produces class probabilities
• Unknown emitters exhibit high distance to all known prototypes

This approach is particularly effective for rapid IoT device onboarding where only 1-5 enrollment transmissions are available per authorized device.

The process of aligning a model's predicted probability of correctness with its actual empirical accuracy, essential for reliable open set rejection thresholds. Key techniques:

• Temperature Scaling: Divides logits by a learned scalar to soften SoftMax outputs
• Platt Scaling: Fits a logistic regression model to the raw logits
• Isotonic Regression: Learns a non-parametric monotonic calibration function

Well-calibrated embeddings ensure that when a model assigns 95% confidence to a known emitter, it is correct 95% of the time. Unknown devices produce uniformly low, well-calibrated probabilities across all known classes.

The reducible model uncertainty arising from a lack of knowledge or data, which is high for inputs far from the training distribution. In feature embeddings:

• Distance to Training Data: Embeddings far from any known class centroid indicate high epistemic uncertainty
• Monte Carlo Dropout: Multiple stochastic forward passes produce variance in embeddings for unknown inputs
• Evidential Deep Learning: Places a Dirichlet distribution over class probabilities to explicitly model evidence

High epistemic uncertainty is the primary signal for triggering open set rejection, distinguishing truly unknown emitters from known devices in noisy conditions.