Enhancing Channel Robustness in LoRa Radio Frequency Fingerprinting through a Physics-Informed Residual Spectrogram †

Radio Frequency Fingerprinting Identification (RFFI) is a lightweight authentication method at the physical layer for LoRa IoT devices that avoids cryptographic overhead. However, current deep learning pipelines for RFFI face a robustness gap in that their identification accuracy can decline in time-varying multipath environments caused by real-world mobility, especially in non-line-of-sight (NLOS) settings. Earlier mitigations rely on data augmentation or implicit channel normalization, yet the resulting features often remain dominated by the deterministic LoRa chirp, which can overshadow device-specific hardware signatures. We proposed a physics-informed residual (PIR) framework that incorporates domain knowledge directly into the extraction of features, thereby enhancing the detection of these hardware imperfections. PIR aligns the received preamble and, in the Channel Independent Spectrogram (CIS) domain, subtracts an equally processed ideal LoRa reference. This operation emphasizes residual patterns linked to transmitter hardware impairments and improves numerical stability. We assess the method on a public dataset of commercial LoRa devices, using a homogeneous set of 45 SX1276-based transceivers split across training, open-set enrollment, and rogue device evaluation. The results indicate more robust performance in most scenarios while maintaining strong open-set rogue detection with Area Under the Curve (AUC) greater than 0.99. These findings suggest that physics-informed representation design can enhance the reliability of LoRa RFFI under dynamic deployment conditions.