One-Shot Enrollment

The core mechanism distills a complete, verifiable identity from one transmission burst. Unlike traditional biometric systems requiring multiple samples for statistical averaging, one-shot enrollment extracts a robust feature vector from a single capture.

• Process: A raw IQ waveform is captured once, preprocessed, and passed through a feature extractor to generate an embedding.
• Challenge: The model must be invariant to channel noise and transient conditions present in that single snapshot.
• Outcome: A stored template vector ready for future cosine similarity comparisons.

One-shot enrollment relies on metric learning frameworks, typically Siamese or Prototypical Networks, to function. The model learns a distance function where genuine and impostor signals are separable.

• Embedding Space: The single enrollment sample is projected into a high-dimensional space where Euclidean or cosine distance defines identity.
• No Retraining: New devices are enrolled by simply storing their embedding; the neural network weights remain frozen.
• Thresholding: Authentication is a simple distance check against a calibrated threshold, enabling sub-millisecond verification.

The primary technical hurdle is ensuring the single enrollment sample is not corrupted by multipath fading or interference. Channel-robust learning techniques are mandatory.

• Domain Adversarial Training: A gradient reversal layer forces the feature extractor to ignore channel-specific artifacts.
• Data Augmentation: The single sample is synthetically distorted with simulated Rayleigh fading and additive white Gaussian noise (AWGN) during training to improve generalization.
• Cyclostationary Anchors: Features tied to the signal's periodic statistical structure remain stable even when the channel varies.

A one-shot enrollment system must inherently support open set recognition. It cannot assume all future devices will be pre-registered.

• Unknown Rejection: If a query embedding's distance to the nearest enrolled template exceeds a calibrated threshold, the device is flagged as unknown.
• Extreme Value Theory (EVT): Statistical models like Weibull distributions are fitted to the tail of the match-score distribution to set rigorous rejection boundaries.
• Novelty Detection: This prevents zero-day rogue transmitters from being incorrectly matched to an existing enrolled identity.

Hardware impairments drift slowly over time due to temperature variation and component aging. A pure one-shot enrollment is a snapshot that can become stale.

• Continuous Authentication: Subsequent successful authentications provide new samples that can update the stored template via exponential moving average.
• Drift Compensation: The system tracks the vector shift of the centroid over time without requiring a full re-enrollment.
• Catastrophic Forgetting Prevention: Elastic Weight Consolidation (EWC) can be applied to the feature extractor if fine-tuning is used to adapt to drift.

A system relying on a single sample is a high-value target for replay attacks and adversarial perturbation. Defensive measures are non-negotiable.

• Liveness Detection: The enrollment process must verify the sample is from a live, active transmission and not a recorded replay.
• Adversarial Training: The feature extractor is hardened against gradient-based perturbations designed to cause misclassification.
• Physical Unclonable Function (PUF) Correlation: The extracted fingerprint should correlate with the underlying hardware PUF, making it mathematically unclonable.

The broader machine learning paradigm that encompasses one-shot enrollment. FSL trains models to generalize from k examples per class, where k is typically between 1 and 5. Unlike traditional deep learning which requires thousands of labeled samples, FSL leverages prior knowledge from related tasks to make accurate predictions on novel classes.

• N-way K-shot classification defines the task format
• Relies heavily on episodic training to simulate deployment conditions
• Critical when capturing multiple enrollment samples is impractical or insecure

A metric-based architecture particularly well-suited for one-shot enrollment. The network learns an embedding function that maps inputs to a vector space where a single support sample defines a class prototype. Classification is performed by finding the nearest prototype.

• Each class is represented by the mean embedding of its support set
• With one-shot, the single sample is the prototype
• Uses Euclidean distance in the learned metric space
• Simple, computationally efficient, and highly effective for RF fingerprinting

A twin-network architecture that learns a similarity function between pairs of inputs rather than explicit class boundaries. During one-shot enrollment, the stored template is compared against a new capture using the learned distance metric.

• Two identical subnetworks with shared weights process both inputs
• Output is a similarity score, not a class prediction
• Naturally suited for open set recognition—rejecting unknown devices
• Training uses contrastive loss or triplet loss to structure the embedding space

The contrastive loss function that powers many one-shot enrollment systems. It operates on triplets of samples: an anchor, a positive (same class), and a negative (different class). The loss forces the anchor-positive distance to be smaller than the anchor-negative distance by at least a margin.

• Formulation: max(d(a,p) - d(a,n) + margin, 0)
• Hard negative mining selects the most challenging negatives for efficient training
• Directly optimizes for the relative ranking needed in one-shot verification

The lower-dimensional continuous vector space where one-shot enrollment templates and query samples are compared. A well-structured embedding space exhibits intra-class compactness and inter-class separability.

• Dimensionality typically ranges from 128 to 512 for RF applications
• Distance metrics: cosine similarity, Euclidean, or Mahalanobis
• Quality directly determines False Acceptance Rate (FAR) and False Rejection Rate (FRR)
• Visualizable with t-SNE or UMAP for debugging enrollment quality

A critical capability for one-shot enrollment systems deployed in the wild. The model must not only match enrolled devices but also reject unknown emitters that were never seen during training or enrollment.

• Requires a confidence threshold on similarity scores
• Out-of-Distribution (OOD) detection mechanisms prevent forced classification
• Essential for detecting spoofing attacks and unauthorized transmitters
• Extreme value theory and Weibull distributions often model the rejection boundary