Calibrated Open-Set Prototypes for Cross-Domain Radio-Frequency Emitter Recognition

Radio-frequency (RF) emitter recognition is a physical-layer tool for wireless security, but recognition accuracy often degrades when the acquisition domain changes. A deployed receiver must also reject unseen emitters and later register novel identities from only a few labeled examples. We address these requirements with a compact prototype framework that uses one shared encoder and two decision-layer modules. Balanced open-set calibration (BOC) maps prototype-score statistics to a calibrated known-class probability and selects a validation-based rejection threshold. Support-aware prototype aggregation (SPA) constructs few-shot novel-class prototypes using centrality weighting, support trimming, and medoid blending. On a curated POWDER cross-day base-station protocol, the full system improves open-set macro-F1 from 55.23% to 69.47%. It also improves 10-shot generalized strict accuracy from 42.33% to 65.54%, relative to a supervised prototype baseline. On a curated Hovering UAV protocol, it improves closed-set accuracy from 33.14% to 59.86%. These gains are obtained without enlarging the deployed neural encoder, supporting RF recognition with calibrated rejection, support-aware registration, and a compact neural footprint.