Open Set Recognition

In congested electromagnetic environments, secondary users must identify spectrum holes without causing interference to primary incumbents. An OSR-based classifier identifies known standardized waveforms (LTE, 5G NR, Wi-Fi) while rejecting unknown or proprietary signals as a distinct 'unknown' class. This prevents a cognitive radio from mistakenly classifying a novel radar pulse as a vacant band, avoiding dangerous collisions. The system uses Entropic Open-Set Loss to force high uncertainty on unfamiliar emissions, triggering a safe fallback mode rather than a catastrophic misclassification.

Signals intelligence (SIGINT) operators face an endless stream of adversarial and civilian emitters. A closed-set classifier fails silently when a new enemy waveform appears. OSR models, using Reciprocal Point Learning, map known threat modulations to compact embedding regions. Any signal falling outside these reciprocal boundaries is flagged as a novel threat emitter for immediate human analysis. This capability is critical for identifying low-probability-of-intercept (LPI) signals and adaptive frequency-hopping patterns that were absent from the training corpus.

National regulatory bodies monitor vast frequency ranges to detect illegal or unlicensed transmissions. OSR automates this by learning the 'normal' RF background of a city. Using Mahalanobis Distance in a deep feature space, the system flags any transmission that deviates statistically from known commercial allocations. This allows an inspector to instantly identify a rogue transmitter using a non-standard modulation scheme, rather than manually scanning waterfalls. The system reduces false positives by calibrating epistemic uncertainty against varying urban noise floors.

Ground stations must autonomously demodulate telemetry from diverse, aging, and often undocumented satellite fleets. An OSR system trained on standard CCSDS protocols can lock onto known downlinks while rejecting unknown interference or spoofing attempts. By applying Temperature Scaling to output logits, the classifier provides a calibrated confidence score. If a signal from a newly launched adversary satellite is received, the low confidence triggers a raw IQ recording for forensic analysis rather than attempting a forced, erroneous demodulation.

Factory floors contain a chaotic mix of Bluetooth, proprietary wireless sensors, and high-noise machinery. An OSR-based diagnostic tool learns the spectral signatures of authorized equipment. When a new jammer or a malfunctioning motor radiates an unknown electromagnetic signature, the system detects it as out-of-distribution using an Autoencoder reconstruction error threshold. This allows maintenance crews to pinpoint novel interference sources without needing a pre-labeled database of every possible fault mode.

Vehicle-to-Everything (V2X) communication relies on standardized IEEE 802.11p or C-V2X waveforms. An OSR module within the vehicle's security gateway verifies that incoming safety messages conform to known modulation profiles. Using Deep Ensembles, the variance of predictions across multiple models is measured. A high variance indicates an unknown or adversarial signal, potentially a spoofing attack using a non-standard waveform, which is immediately rejected to prevent the autonomous system from acting on malicious data.

The foundational statistical framework for open set recognition. Extreme Value Theory models the tail behavior of distributions, fitting a Weibull distribution to the distance of correct classifications from their class mean. The OpenMax layer replaces standard SoftMax by recalibrating activation vectors using this theory, explicitly estimating the probability that an input belongs to an unknown class rather than forcing it into a known category.

Specialized training objectives that structure the feature space for rejection. Objectosphere Loss creates a distinct separation by maximizing the feature norm for known samples while minimizing it for unknowns, creating a thresholdable rejection space. Entropic Open-Set Loss forces the network to produce high-entropy, uniform probability distributions for unknown samples, making them easily separable from the low-entropy predictions of known classes.

A classification strategy that represents each known class by a reciprocal point in the embedding space rather than a centroid. Unknown samples are identified by measuring the maximum distance to these reciprocal points. This approach avoids the feature collapse failure mode where all inputs, including unknowns, map to a restricted region, destroying the model's ability to separate known from novel signal types.

Epistemic uncertainty is the model's uncertainty arising from a lack of knowledge, which is reducible with more data and is a key signal for detecting unknown modulation classes. Evidence Deep Learning treats predictions as subjective opinions by placing a Dirichlet distribution over class probabilities, enabling direct quantification of predictive uncertainty for novelty detection without requiring auxiliary outlier datasets.

Practical techniques for improving open set performance. Outlier Exposure regularizes the model by training with an auxiliary dataset of diverse outlier examples, forcing the network to learn a tighter decision boundary. Deep Ensembles train multiple networks with different initializations and use the variance of their predictions as a robust, threshold-independent signal for detecting unknown inputs.

A continuous learning paradigm extending beyond simple detection. An Open World Learner must not only detect unknown classes but also incrementally learn to recognize them when labeled data becomes available, all without suffering from catastrophic forgetting of previous modulation types. This contrasts with static open set recognition by closing the loop on novel signal intelligence.