A Lightweight Open-Set Specific Emitter Identification Based on Broad Learning

Specific Emitter Identification (SEI) exploits subtle, device-specific imperfections in physical-layer signals—known as Radio Frequency Fingerprints (RFFs)—to distinguish authorized transmitters from unauthorized ones. Conventional deep-learning SEI approaches, however, typically require large labeled datasets, incur long training times, consume substantial computational resources, and provide limited interpretability—constraints that hinder deployment on resource-constrained receivers. We propose BLi2, a lightweight Broad Learning (BL) framework engineered to mitigate these challenges for open-set SEI in IoT settings. BLi2 uses an Instantaneous Autocorrelation Representation (IAR) to extract compact, phase-based fingerprints from raw I/Q samples and employs a two-stage open-set recognition pipeline to detect unauthorized emitters before closed-set classification. Experiments on two real-world ADS-B datasets and a Wi-Fi dataset demonstrate that BLi2: (1) attains >98.80% closed-set accuracy for 50 classes (98.72% for 150 classes while using only ≈0.195M parameters—over a half as comparable BL methods) and ≈90% open-set accuracy with 50 unknowns; (2) trains in under 20 s on a Raspberry Pi, maintaining >96.33% accuracy across SNRs; and (3) improves accuracy by >12% compared with raw I/Q baselines in the reported ablation. These results validate that BLi2 provides high identification accuracy while maintaining the computational efficiency required for edge-IoT deployments.