Metric Learning

Unlike traditional classification which learns decision boundaries, metric learning optimizes a distance function directly. The goal is to ensure that the computed distance—often Euclidean or cosine similarity—between semantically similar inputs is minimized, while the distance between dissimilar inputs is maximized. This is typically enforced through contrastive loss or triplet loss functions that operate on pairs or triplets of data points, penalizing the model when relative distances violate a defined margin.

Training relies on the strategic sampling of data tuples rather than individual samples. Key formations include:

• Positive Pairs: Two samples from the same class (e.g., two RF bursts from the same transmitter).
• Negative Pairs: Samples from different classes.
• Triplets: An anchor, a positive match, and a negative match. The selection of hard negatives—dissimilar samples that are currently close in the embedding space—is critical for fast convergence and robust feature learning.

The output of a metric learning model is a mapping to a high-dimensional embedding space where distances are semantically meaningful. This space is structured such that:

• Clusters form for each distinct identity or class.
• Intra-class variance is minimized, pulling all instances of a specific emitter into a tight sphere.
• Inter-class separation is maximized, pushing different identities apart. This geometry enables efficient nearest-neighbor search for identification tasks, even for classes never seen during training.

A defining advantage of metric learning is its ability to perform open set recognition and few-shot generalization. Because the model learns a universal similarity function rather than a fixed set of class weights, it can compare a query sample against a support set of entirely new, previously unseen classes at inference time. This is foundational for few-shot device enrollment, where a new transmitter can be registered with only one or a few examples without retraining the entire network.

To avoid the combinatorial explosion of pairwise comparisons in large datasets, modern metric learning often employs proxy-based losses such as Proxy-NCA or Proxy Anchor Loss. Instead of comparing every sample against all others, each class is represented by a learnable proxy vector. The model only compares data embeddings to these limited proxies, drastically reducing training complexity from O(N²) to O(N) while maintaining high retrieval accuracy and tight clustering behavior.

Advanced metric learning objectives incorporate additive angular margin penalties to enforce stricter intra-class compactness. Techniques like ArcFace or CosFace normalize both the feature embeddings and the class proxy weights, then add a multiplicative or additive angular margin to the target class logit. This forces the model to produce embeddings that are not only closer to their correct proxy but also separated by a precise angular gap, significantly hardening the decision boundary against impostor samples.

A contrastive loss function that structures the embedding space by enforcing a margin-based relationship between three samples:

• Anchor: A reference sample from a specific class
• Positive: Another sample from the same class as the anchor
• Negative: A sample from a different class

The loss minimizes the distance between the anchor and positive while ensuring the negative is at least a margin α farther away than the positive. Formally: L = max(d(a,p) - d(a,n) + α, 0). This directly trains the network to create clusters where intra-class distances are small and inter-class distances are large, making it ideal for device fingerprinting where each transmitter forms its own class.

A metric-based few-shot learning architecture that represents each class by a single prototype vector: the mean of its support set embeddings in the learned metric space.

Classification proceeds by:

1. Embedding all support samples through the network
2. Computing the centroid (prototype) for each class
3. Classifying query samples by finding the nearest prototype via Euclidean distance

For device enrollment, a new transmitter's prototype can be computed from as few as 1-5 RF fingerprint samples, enabling rapid onboarding without retraining the entire network. The non-parametric nature of the classifier means new device classes can be added instantly.

A twin-network architecture composed of two identical subnetworks that share weights and process pairs of inputs in parallel to determine their similarity.

Key characteristics:

• Both branches apply the exact same transformation to their respective inputs
• The outputs are two embedding vectors compared via a distance function (e.g., L1, cosine)
• The final layer outputs a similarity score between 0 and 1
• Trained with contrastive loss or binary cross-entropy on paired data

In RF fingerprinting, Siamese networks excel at verification tasks: given two signal samples, determine if they originated from the same physical transmitter. This pairwise approach is data-efficient and naturally handles open-set scenarios.

A lower-dimensional continuous vector space where semantically similar data points are mapped close together, enabling distance-based comparison and clustering.

Critical properties for device authentication:

• Dimensionality: Typically 64-512 dimensions, balancing expressiveness with computational efficiency
• Compactness: Intra-class variance is minimized; all samples from one transmitter form a tight cluster
• Separability: Inter-class distance is maximized; different transmitters occupy distinct regions
• Smoothness: Small perturbations in input space produce small displacements in embedding space

The embedding space serves as the mathematical foundation for all distance-based authentication decisions. Its quality directly determines False Acceptance Rate (FAR) and False Rejection Rate (FRR).

A normalized similarity measure between two non-zero vectors that calculates the cosine of the angle between them, ranging from -1 (opposite) to 1 (identical).

Formula: cos(θ) = (A·B) / (||A|| × ||B||)

Advantages over Euclidean distance for RF fingerprinting:

• Magnitude-invariant: Focuses on directional similarity, ignoring vector length variations caused by signal power differences
• Bounded range: Outputs are naturally normalized to [-1, 1], simplifying threshold selection
• Efficient computation: Requires only dot product and L2 norm operations

Often used as the final comparison metric in Siamese networks and as an alternative to Euclidean distance in Prototypical Networks when signal amplitude varies significantly across capture sessions.

A meta-learning training strategy where each iteration simulates a complete few-shot task by sampling a small support set and query set from the overall dataset.

Each episode:

1. Randomly selects N classes from the training set
2. Samples K support examples and M query examples per class (an N-way K-shot task)
3. The model learns to classify the query set using only the support set
4. Loss is computed on query predictions and backpropagated

This train-to-learn paradigm explicitly optimizes for rapid adaptation rather than memorization. For device enrollment, episodes mirror real deployment: the model sees a few examples of new transmitters and must immediately authenticate them, matching the exact inference-time distribution.