Epistemic Uncertainty

This uncertainty is inversely proportional to the density of training data in the feature space. In regions where the model has observed many examples, epistemic uncertainty collapses to near zero. In contrast, for out-of-distribution (OOD) inputs or novel emitter types far from known clusters, the model's lack of knowledge causes a sharp spike in uncertainty. This property makes it a direct measure of model ignorance rather than data noise.

Unlike aleatoric uncertainty, which stems from irreducible sensor noise, epistemic uncertainty can be systematically reduced. By retraining the model on the specific inputs that triggered high uncertainty, the knowledge gap is closed. This drives active learning loops in spectrum surveillance, where a system autonomously queries a human analyst for labels only on the most uncertain, previously unseen emitters.

Standard neural networks provide overconfident point estimates. Epistemic uncertainty is quantified by placing distributions over model weights. Practical approximations include:

• Monte Carlo Dropout: Applying dropout at inference to generate stochastic forward passes.
• Deep Ensembles: Training multiple models with different initializations and treating their disagreement as uncertainty. The variance of predictions across these passes defines the model's confidence in its own knowledge.

In open set emitter recognition, the model must reject unknown devices rather than forcibly misclassifying them. Epistemic uncertainty provides the rejection logic: if the predictive variance exceeds a calibrated threshold, the input is flagged as unknown. This prevents a rogue transmitter from being silently classified as an authorized device, a critical security feature for physical layer authentication.

Geometrically, epistemic uncertainty correlates with the distance from the learned data manifold. Techniques like Mahalanobis distance or feature embedding norms measure how far a test sample lies from the convex hull of known classes. A large distance implies the model is extrapolating rather than interpolating, signaling a high probability that the input belongs to an unseen emitter category.

This approach places a Dirichlet distribution directly over the class probabilities, bypassing the need for sampling. The model outputs evidence for each class, and the total evidence mass inversely relates to epistemic uncertainty. When an input receives uniformly low evidence across all known classes, the model expresses high uncertainty, indicating a lack of supporting knowledge for any known emitter type.

The irreducible statistical noise inherent in the data itself, such as sensor thermal noise, multipath fading, or overlapping signal constellations. Unlike epistemic uncertainty, collecting more training data cannot reduce aleatoric uncertainty. It is typically modeled by predicting a variance parameter alongside the mean, allowing the model to express 'I know this measurement is noisy.' In RF fingerprinting, high aleatoric uncertainty often indicates a low-SNR environment rather than an unknown device.

A practical Bayesian approximation that applies dropout at inference time to generate multiple stochastic forward passes through the network. The variance across these predictions serves as a proxy for epistemic uncertainty:

• High variance across passes suggests the input is far from training data (unknown emitter)
• Low variance indicates a confident, in-distribution classification This technique requires no architectural changes, making it a lightweight method for adding uncertainty estimation to existing fingerprinting models.

A method that places a Dirichlet distribution over class probabilities rather than outputting a single point estimate. The model learns to predict evidence for each class, from which belief masses and uncertainty are derived:

• High evidence + low uncertainty: Confident known-class prediction
• Low evidence + high uncertainty: Unknown emitter detected This framework naturally decomposes total uncertainty into epistemic and aleatoric components, providing a mathematically grounded rejection mechanism for open-set emitter recognition.

A distribution-free statistical framework that produces prediction sets with a guaranteed marginal coverage probability (e.g., 95%). For open-set recognition:

• Singleton sets indicate confident known-class identification
• Empty sets signal an unknown emitter, triggering rejection
• Large sets reflect high epistemic uncertainty Conformal prediction provides rigorous, finite-sample validity guarantees without assumptions about the underlying data distribution, making it valuable for safety-critical spectrum surveillance applications.

A framework that learns an energy function mapping inputs to a scalar score:

• Low energy: In-distribution, known emitter classes
• High energy: Out-of-distribution, unknown devices EBMs are trained to assign lower energy to training data and higher energy elsewhere. During inference, an energy threshold serves as the rejection boundary. Unlike SoftMax classifiers, EBMs do not suffer from overconfidence on unknowns, as the energy score is unbounded and can rise arbitrarily high for novel inputs far from the training manifold.

A post-hoc confidence calibration technique that divides the logits by a learned scalar parameter T (temperature) before applying SoftMax:

• T > 1: Softens probabilities, reducing overconfidence
• T < 1: Sharpens probabilities
• T = 1: Standard SoftMax Proper calibration is critical for open-set rejection because raw neural network outputs are typically miscalibrated and overconfident. A well-calibrated model's confidence scores can be directly thresholded to separate known emitters from unknown ones.