Open Set Recognition

Traditional closed-set classification assumes all test classes are present during training, forcing a prediction into one of K known categories—even for an unknown emitter. Open Set Recognition relaxes this assumption, introducing a (K+1)-th implicit class for 'unknown.' The model must operate with classifiable knowns, detected unknowns, and an open space risk—the theoretical risk of labeling an unknown sample as known. This is formalized by the compact abating probability model, where known class probabilities must diminish to zero far from training data.

OpenMax is the foundational algorithm that adapts a standard neural network for open set recognition without retraining. It replaces the final SoftMax layer with a calibrated rejection mechanism:

• It fits a Weibull distribution to the penultimate layer's activation vectors (the Meta-Recognition step) using Extreme Value Theory (EVT).
• For each class, it estimates the probability of an input being an outlier based on its distance from the class mean activation vector.
• It recalibrates scores and explicitly reserves a probability mass for the 'unknown' class, rejecting inputs that fall in the open space between known clusters.

This is the core theoretical property required for a robust OSR solution. A classifier exhibits compact abating probability if:

• The probability of a known class is strictly bounded within a compact region of the feature space.
• Beyond that region, the probability abates (diminishes) to zero.
• This ensures that open space risk—the relative measure of space labeled as 'known' compared to the space labeled as 'unknown'—is minimized. Without this property, a model will confidently misclassify unknown emitters into known categories.

Extreme Value Theory (EVT) provides the statistical foundation for calibrating the boundary between known and unknown. Instead of modeling the entire distribution of activation scores, EVT focuses on the tail of the distribution—the extreme values. By fitting a Weibull, Fréchet, or Gumbel distribution to the distances between correctly classified samples and their class mean, the model can compute a calibrated probability that a new, distant sample belongs to an entirely new class. This is the Meta-Recognition problem: recognizing when the model doesn't know.

An alternative to EVT-based methods uses generative models to detect unknowns. A Variational Autoencoder (VAE) or Generative Adversarial Network (GAN) is trained exclusively on known emitter signatures. During inference:

• A known input will be reconstructed with low error.
• An unknown emitter will produce a high reconstruction error because its features fall outside the learned manifold.
• This error signal serves as a novelty score. A threshold is set; inputs exceeding it are rejected as open set. This method directly models the support of the known data distribution.

Open Space Risk is the theoretical measure of a model's vulnerability to labeling unknown space as known. Formally, it is the ratio of the measure of positively labeled open space to the measure of positively labeled known data. Minimizing this risk is the objective of OSR. Practical strategies include:

• G-OpenMax: Uses generative models to synthesize open-set samples for training.
• CROSR: Jointly trains a classifier and a reconstructor to learn a latent representation where known classes are compact and separable.
• Prototype Learning: Forces features to cluster tightly around learned class prototypes, leaving vast regions of the latent space empty for rejection.

In electronic warfare and spectrum enforcement, OSR models continuously monitor the electromagnetic environment to identify known friendly or licensed transmitters while flagging any unknown or unauthorized emissions. Unlike closed-set classifiers that forcibly map a novel adversary signal to a known class, an OSR system triggers an alert for human analysis or autonomous countermeasures. This capability is critical for identifying pirate radio stations, unregistered drone controllers, or hostile jammers operating on unexpected frequencies.

Conventional intrusion detection systems rely on signature databases and fail against zero-day exploits. OSR-enabled network monitors learn the baseline behavior of authorized IoT devices through their RF fingerprints or traffic patterns. When a spoofed device or malware-compromised node exhibits behavioral features outside the known distribution, the OSR model rejects it as unknown, blocking the attack before a signature is ever created. This provides protocol-agnostic authentication at the physical layer.

Radio telescopes generate massive volumes of spectral data, where the goal is to identify novel astrophysical phenomena amidst known interference. OSR models are trained on labeled examples of terrestrial RFI, satellite passes, and known pulsar signatures. The system then rejects these known classes to isolate unexplained transient signals—potential candidates for new astronomical discoveries or even technosignatures. This automates what was previously a manual, labor-intensive search for the unknown.

Electronics supply chains are vulnerable to sophisticated counterfeit ICs that pass basic functional tests. OSR models are trained on the RF or power-side-channel fingerprints of authentic components from trusted foundries. During incoming inspection, a counterfeit chip—even a functionally identical clone—will exhibit subtle hardware impairments that place it outside the known-authentic feature distribution. The OSR system rejects the component as unknown, preventing it from entering mission-critical assemblies.

Voice biometric systems must not only verify enrolled users but also reject impostors and synthetic voices never seen during training. OSR architectures learn a compact embedding space for authorized speakers and calibrate a rejection boundary using Extreme Value Theory. When a deepfake audio sample or an unenrolled speaker attempts access, the model recognizes the voice as outside the known distribution and denies authentication, providing robustness against AI-generated voice attacks.

Autonomous vehicles and robots operating in unstructured environments inevitably encounter objects and scenarios absent from training data. An OSR-based perception stack identifies these unknowns—such as an unusual obstacle or a novel road sign—and triggers a conservative safety policy (e.g., slow down, increase following distance) rather than misclassifying the object and proceeding dangerously. This bridges the gap between brittle closed-set perception and the open-world reality of physical deployment.