Prototypical Network

Each known emitter class is represented by a single prototype vector—the mean of its support set embeddings in the learned metric space. For a class k with support examples S_k, the prototype is:

• c_k = (1/|S_k|) Σ f_φ(x_i) for all x_i in S_k
• The embedding function f_φ is typically a deep convolutional neural network
• Averaging provides natural noise resilience, smoothing out individual signal anomalies
• This simple mean operation requires no additional learnable parameters, preventing overfitting in few-shot regimes

A query signal is classified by computing its distance to every prototype and selecting the nearest class. The network is trained to minimize the negative log-probability of the correct class via a softmax over distances:

• p(y=k|x) ∝ exp(-d(f_φ(x), c_k))
• Euclidean distance corresponds to a Gaussian mixture model assumption in the embedding space
• Bregman divergences generalize the framework—any regular exponential family distribution has a corresponding divergence
• Squared Euclidean distance is the default, providing closed-form prototype updates and efficient computation

Training mimics the few-shot inference scenario by constructing episodes—mini-batches structured as N-way K-shot classification problems:

• Each episode samples N random emitter classes from the training set
• K support examples per class form the prototype, with query examples used for loss computation
• This meta-learning strategy ensures the model generalizes to unseen emitter classes at test time
• Higher N (e.g., 20-way) during training improves the learned metric space's discriminative capacity

Prototypical Networks naturally support open-set emitter identification—detecting rogue devices not present in the support set:

• A query is rejected if its minimum distance to any prototype exceeds a calibrated threshold τ
• The threshold is tuned on a validation set to balance false acceptance rate (FAR) and false rejection rate (FRR)
• This provides inherent novelty detection without architectural modification
• Distance magnitudes can be interpreted as confidence scores for mission-critical authentication decisions

RF fingerprints captured under different channel conditions can cause prototype drift. Domain-adversarial training is integrated to enforce channel invariance:

• A gradient reversal layer between the embedding network and a domain classifier forces channel-agnostic features
• Prototypes computed from support examples in one channel condition remain valid for queries in another
• This addresses the critical channel robustness problem in real-world SEI deployments
• Combined with episodic training, the model learns to isolate hardware impairments from propagation effects

New emitters can be enrolled into the system without any gradient updates or access to previous training data:

• Capture K support examples of the new device and compute its prototype via forward pass
• Append the prototype to the existing prototype bank—no fine-tuning required
• This enables zero-cost enrollment in dynamic environments where new devices appear continuously
• Prototype aging can be managed by maintaining a rolling buffer of recent support examples to track device aging drift

A classification paradigm where a model learns a distance function over embeddings rather than explicit class boundaries. Prototypical Networks are a canonical example, computing Euclidean distance to class prototypes (mean vectors). This contrasts with optimization-based methods like MAML. In RF fingerprinting, metric learning is preferred because the embedding space naturally clusters signals by hardware impairment similarity, enabling open-set recognition of unknown emitters.

The two data partitions defining an episode during few-shot training:

• Support Set: A small labeled sample (e.g., 5 examples per emitter) used to compute class prototypes
• Query Set: Unlabeled samples used to evaluate classification accuracy against the prototypes

This episodic training mimics deployment conditions where only a handful of reference signals exist for each known transmitter. The model learns to generalize across many such episodes rather than memorizing fixed classes.

The neural network that maps raw IQ samples or RF fingerprints into a high-dimensional embedding space where Euclidean distance corresponds to emitter similarity. Common architectures include:

• 1D Convolutional Networks: For raw IQ time-series data
• Wavelet Scattering Networks: For translation-invariant RF representations
• Siamese-style encoders: Pre-trained with contrastive loss

The embedding network is trained end-to-end with the prototype loss, learning to extract features that minimize intra-class distance while maximizing inter-class separation.

The class prototype is the mean vector of all support set embeddings for a given emitter:

c_k = (1/|S_k|) * Σ f_φ(x_i)

where f_φ is the embedding network and S_k is the support set for class k. This simple averaging provides noise robustness—individual corrupted samples are averaged out. For RF applications with channel variation, weighted averaging or attention mechanisms can emphasize cleaner, higher-SNR support examples when computing the prototype.

A query sample x is classified by computing its softmax over distances to all prototypes:

p(y=k|x) ∝ exp(-d(f_φ(x), c_k))

The negative Euclidean distance ensures closer prototypes yield higher probabilities. This inherently supports open-set recognition: if the minimum distance exceeds a calibrated threshold, the query is rejected as an unknown emitter. This is critical for rogue device detection in physical layer security.