Radio-frequency fingerprinting (RFF) provides physical-layer device authentication for IoT networks, but standard deep-learning clas- sifiers collapse under noisy and fading channels (e.g., from 91.8% clean accuracy to 0.9% at −10 dB SNR). I propose a noise-aware meta-learning framework that combines mixed-channel episodic train- ing with SNR-conditioned prototype weighting. During meta-training, samples undergo both AWGN and Rician fading, forcing the embed- ding network to learn channel-invariant representations. At inference, a softmax-weighted prototype mechanism prioritises high-SNR support samples, improving robustness. Evaluated on the SMoRFFI dataset (123 IEEE 802.11g devices, 122,511 IQ segments), SNR-ProtoNet achieves 32.6% accuracy at −10 dB under AWGN and 31.0% under Rician fading, representing a 35-fold improvement over the baseline CNN. Under Rician fading, the SNR-conditioned model consistently outperforms the noise-augmented ProtoNet, with gains up to +7.7 percentage points. The architecture adds only 20,672 parameters to the standard ProtoNet backbone, maintaining suitability for resource- constrained IoT devices.
Bharat Paudel (Wed,) studied this question.