Similarity-Adaptive Framework for Semi-Supervised Open-World Specific Emitter Identification

Specific emitter identification (SEI) is a physical-layer authentication technique that identifies devices by extracting radio frequency fingerprints (RFFs) from received signals. Open-set SEI (OS-SEI) refers to classifying known classes while rejecting unknown classes, which typically requires a sufficient amount of labeled training samples. However, in open-world scenarios, labeled samples are often limited, and unlabeled samples may contain unknown classes. Moreover, open-world recognition not only requires detecting unknown class samples but also identifying specific novel classes within these unknown samples and integrating them into the recognition model. Current OS-SEI methods can only categorize all unknown samples as a single class, lacking the ability to further differentiate these unknown classes. To address these challenges, we formulate a novel semi-supervised open-world SEI (SSOW-SEI) problem, which aims to overcome the shortcomings of OS-SEI in utilizing unlabeled data, distinguishing unknown classes, and addressing class distribution mismatches between labeled and unlabeled data. Furthermore, we develop an end-to-end similarity-adaptive (SAA) framework for SSOW-SEI. Specifically, after automatically extracting sample features, SAA first identifies novel classes by measuring pairwise similarities between the features, and then recognizes known classes using adaptive cross-entropy, which balances the learning rate between known and novel classes to prevent model bias toward known classes. Additionally, entropy regularization is applied to mitigate model overfitting. Extensive experimental results demonstrate that the proposed SAA framework effectively leverages limited labeled data, handles large volumes of unlabeled data, and accurately identifies both known and novel classes. The results also highlight its strong generalization, stability, and enhanced adaptability to novel classes.