From Uncertainty to Inconsistency: Open-Set RF Fingerprint Identification

ICLR 2026 Conference Submission25310 Authors

Keywords: Open-set recognition, radio frequency fingerprint identification, deep learning

TL;DR: Inspired by an interesting observation that predictions for unknown classes across multiple models exhibit high inconsistency, while predictions for known classes show high consistency, we propose an inconsistency based open-set RFFI approach.

Abstract: The rejection of unknown devices outside the known categories is crucial for radio frequency fingerprint identification (RFFI). Current open-set recognition (OSR) methods rely on the uncertainty of the model output, where unknown classes exhibit low confidence and vice versa for known classes. However, we demonstrate that uncertainty-based methods face a significant challenge, particularly in RFFI, which is termed ‘‘Overconfidence on Unknown Signal Segments’’ (OUSS), where unknown signal segments are misclassified with high confidence, directly contradicting the expected low-confidence characteristic for unknown classes. Inspired by an interesting observation that predictions for unknown classes across multiple models exhibit high inconsistency, while known classes exhibit the opposite, we propose to leverage decision entropy to quantify the inconsistency. Based on the decision entropy, we propose an inconsistency based open-set RFFI approach (IncOS-RFFI). We conduct extensive experiments on the seven open-source radio frequency fingerprint datasets with seventeen benchmarks and demonstrate the effectiveness of our proposed IncOS-RFFI compared to existing OSR algorithms.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 25310