A Multigranularity Embedding Guided Open-Set Recognition for Fine-Grained Specific Emitter Identification

Abstract

In wireless communication security, specific emitter identification (SEI) is a fundamental mechanism for device authentication and access regulation. Under open-set conditions, the model must discriminate devices observed during training and detect and reject previously unknown emitters, which may correspond to unauthorized or unidentified entities. However, existing methods for open-set SEI (OS-SEI) frequently require extensive annotated samples from known classes and finely tuned rejection thresholds, which restrict their ability to partition unknown categories and limit their robustness under non-stationary operational environments. To address these limitations, this paper proposes multigranularity embedding guided open recognition (MGEGOR), a prototype-based open-set SEI framework inspired by generalized zero-shot learning. The proposed framework integrates contrastive representation learning, prototype-based embedding, and open-set generalization, and it employs multiscale feature extraction with attention alignment to construct a discriminative embedding space where known and unknown categories can be modeled jointly. Under the open-set formulation, the proposed MGEGOR achieves significant accuracy improvements over state-of-the-art baseline methods on the LoRa_RFFI dataset and measured datasets, demonstrating heterogeneous open-environment robustness and effective unknown device identification.