An Open-Set Few-Shot Class Incremental Learning Framework for Specific Emitter Identification

Specific Emitter Identification (SEI) technology enhances the physical-layer security of the Internet of Things (IoT). However, traditional SEI systems primarily focus on identification tasks within static closed-set environments, limiting their applicability in dynamic open-set scenarios. To address the challenges of Few-Shot Incremental Open-Set Recognition (FIOSR), which involves the expansion of specific emitter categories and Open-Set Recognition (OSR), we propose an Open-Set Few-Shot Class Incremental Learning (OFSCIL) framework. Our framework decouples the feature extractor from the classifier, and optimizes the extractor using an innovative integrated loss function and class augmentation. Moreover, we design a sustainably evolving prototype classifier with a prototype calibration method to alleviate the catastrophic forgetting and overfitting in the Few-Shot Class Incremental Learning (FSCIL) phase. For recognizing unknown categories, we design an open-set classifier based on prototype distance. Additionally, a semi-supervised clustering algorithm using cosine distance is introduced to further distinguish unknown signals. Experimental results on the benchmark dataset show that OFSCIL is well adapted to OSR, FSCIL and FIOSR tasks, highlighting its capability to balance computational efficiency and accuracy.