Open-Set Recognition for RF

The core mathematical challenge is managing open-space risk—the probability that an unknown sample is incorrectly assigned a high confidence score for a known class. This is not merely threshold tuning; it requires the model to learn a compact abatement boundary around each known emitter's feature manifold.

• Extreme Value Theory (EVT) is often applied to model the tail of the distance distribution, calibrating rejection thresholds based on Weibull distributions rather than arbitrary heuristics.
• The system must bound open-space risk while maintaining high closed-set accuracy, a fundamental trade-off formalized in the open-set risk function.

A specialized deep learning paradigm where each known class is assigned a reciprocal point—a learned representation of the latent space that is maximally distant from that class's embeddings. During inference, the distance to all reciprocal points is measured.

• If a sample is far from all reciprocal points, it is classified as the corresponding known class.
• If a sample is close to every reciprocal point, it is rejected as unknown. This creates a geometrically principled decision boundary without requiring unknown samples during training.

Since real rogue emitter data is scarce, models are hardened using synthetic unknown generation. Generative Adversarial Networks (GANs) or diffusion models synthesize plausible yet novel RF signatures that lie near the boundary of known class manifolds.

• These generated samples act as proxy unknowns, forcing the classifier to tighten its decision boundaries.
• The generator is trained to produce signals that are confusing—close to known classes but statistically distinct—creating a robust discriminative boundary without prior knowledge of actual threats.

A non-parametric open-set classifier that fits a Weibull distribution to the distance between a point and its nearest neighbors within each class. The EVM does not learn a global decision hyperplane; it performs local probabilistic calibration around each training sample.

• The Weibull CDF gives the probability that a query point belongs to the same class as its nearest reference point.
• A sample is rejected as unknown if its probability of inclusion falls below a calibrated threshold for all known classes. This provides a statistically rigorous rejection mechanism.

Adapted from face recognition, this approach modifies the softmax loss to enforce an additive angular margin between known class prototypes in hyperspherical embedding space. The resulting embeddings are compact within classes and separated by enforced angular gaps.

• The ArcFace or CosFace loss functions push known class weights apart, creating a void in the embedding space.
• Unknown samples naturally fall into these angular gaps, making them detectable via cosine similarity thresholds without explicit unknown training data.

Standard softmax confidence is dangerously miscalibrated for open-set RF—unknown inputs can produce high-confidence predictions for known classes. Temperature scaling and energy-based models are deployed to recalibrate output probabilities.

• The energy score, derived from the log-sum-exp of logits, provides a density-based measure that is lower for in-distribution samples and higher for unknowns.
• Monte Carlo Dropout at inference time estimates predictive uncertainty; high variance across stochastic forward passes signals an out-of-distribution input, triggering rejection.

In contested electromagnetic environments, open-set models identify friendly emitters via Specific Emitter Identification (SEI) while flagging unknown adversarial signals. This prevents reliance on a closed-set library that fails against novel jammers or spoofed waveforms.

• Threat Library Agnosticism: Detects zero-day electronic attacks without prior signature.
• I/Q-Level Analysis: Operates on raw in-phase and quadrature samples to catch subtle hardware anomalies.
• Real-Time Triage: Automatically rejects unknown signals for further SIGINT analysis.

Enterprise IoT networks use open-set recognition to enforce Physical-Layer Authentication during device provisioning. Known, authorized transmitters are admitted based on their RF-DNA, while rogue devices attempting MAC Address Spoofing are rejected outright.

• Continuous Authentication: Persistently validates physical-layer identity, not just at login.
• Clone Rejection: Detects sophisticated hardware clones via non-cloneable RF PUF characteristics.
• Protocol Agnostic: Secures Zigbee, LoRaWAN, and BLE devices without modifying upper-layer stacks.

Power grids and water treatment facilities deploy open-set RF monitors to detect unauthorized transmitters near SCADA wireless telemetry links. The system learns the Cyclostationary Feature signatures of legitimate field devices and alarms on any unknown emitter.

• Anomaly Detection: Identifies reconnaissance or jamming attempts without prior attack signatures.
• Geospatial Correlation: Integrates with Radio Environment Maps to localize rogue sources.
• False Positive Mitigation: Open-set thresholds are tuned to maintain operational continuity while ensuring security.

National regulatory authorities use open-set classifiers in mobile monitoring stations to detect unlicensed or out-of-band transmissions. The model identifies known compliant emitters and isolates unknown signals for enforcement action.

• Wideband Monitoring: Processes gigahertz of spectrum simultaneously using Wideband Signal Processing.
• Automatic Modulation Classification: Identifies the transmission scheme of unknown signals as a first forensic step.
• Evidence Generation: Produces time-stamped, geotagged logs of unknown emitter activity for legal proceedings.

Connected vehicles use open-set recognition to validate Vehicle-to-Everything (V2X) messages at the physical layer. A vehicle's onboard unit verifies that received Basic Safety Messages originate from known, trusted transmitters and rejects spoofed messages from unknown sources.

• Misbehavior Detection: Catches Sybil attacks where a single attacker impersonates multiple phantom vehicles.
• Channel-Robust Operation: Maintains accuracy despite high Doppler shift and multipath fading.
• Edge Deployment: Runs optimized inference on automotive-grade Neural Processing Units.

Airspace security systems use open-set recognition to distinguish authorized commercial drones from unknown or hostile unmanned aerial systems (UAS). The model fingerprints the Turn-On Transient and Preamble Distortion of drone remote control signals.

• Open-Set Drone ID: Differentiates between a known fleet operator's drone and an unknown intruder.
• Protocol-Agnostic Detection: Works across DJI, Autel, and custom flight controllers.
• Integration with Radar: Fuses RF classification with radar tracks for high-confidence threat assessment.