Operational RF classifiers encounter signals outside
the closed-set label taxonomy. We treat Unknown as a first-class
outcome via score thresholding on logits-derived confidence (maxprobability), predictive entropy, and energy score. We evaluate
with OSCR curves (Correct Classification Rate vs. Unknown
False-Positive Rate) and AU-PR for unknown detection. With
lightweight hooks in the ensemble probability path, we realize
robust abstention with <0.1 ms overhead and maintain utility
(accuracy×coverage) across SNR bins.
Open-set recognition (OSR) addresses the mismatch between training labels and real-world observations. In RFML
pipelines, abstention is often treated as a failure; we instead
elevate Unknown to a policy outcome with measurable tradeoffs. We analyze simple gates on: (i) max softmax, (ii)
predictive entropy, and (iii) logit-energy, and report OSCR [1],
[2] and AU-PR(Unknown).
Related Papers