A Model Compression Framework for Edge-Deployable LoRa Device Authentication

The widespread deployment of LoRa in IoT networks necessitates efficient physical-layer authentication to address security vulnerabilities introduced by unauthorized devices. While deep learning-based RF fingerprinting offers promising solutions, practical implementation faces challenges in computational complexity and channel variation sensitivity. This paper presents a model compression framework for LoRa device identification that combines knowledge distillation, unstructured pruning and adaptive quantization. The proposed method reduces model parameters more than $70 \%$ (5.88 MB final size) while maintaining 95.5% accuracy. Experimental validation demonstrates $\mathbf{0. 9 3 9}$ anomaly detection AUC across commercial LoRa devices at SNR $\geq 20 \mathrm{~dB}$. Implementation on Raspberry Pi 4B achieves 7-10s inference latency, confirming practical deployability. These results suggest an effective balance between security and efficiency for IoT edge applications.