SNR-ProtoNet: Robust Few-Shot RF Fingerprinting via SNR-Conditioned Prototypical Networks

Radio-frequency fingerprinting (RFF) provides physical-layer device authentication for Internet of Things (IoT) networks, but standard deep-learning classifiers collapse under noisy and fading channels, dropping from 91.8% clean-signal accuracy to 0.9% at −10 dB SNR under AWGN. SNR-ProtoNet combines mixed-channel episodic training with SNR-conditioned prototype weighting, targeting robustness to amplitude-dominated channel degradation (additive noise and flat Rician fading) in same-model, single-receiver IoT deployments. Training exposes samples to AWGN and Rician fading across a wide SNR range, learning channel-invariant features; a softmax-weighted prototype mechanism prioritizes high-SNR support samples at inference. Evaluated on the public SMoRFFI dataset (123 same-model IEEE 802.11g devices, 122,511 IQ segments) under a cross-channel protocol with clean enrollment and degraded queries, SNR-ProtoNet achieves 32.6% identification accuracy at −10 dB under AWGN — a 36-fold improvement over a conventional CNN — and consistently outperforms channel-diverse training alone under Rician fading, with gains of 2.9 to 7.7 percentage points, while adding only 20,672 parameters (0.6% additional inference-time FLOPs) over a standard Prototypical Network backbone. Beyond these core results, this work provides matched-SNR baseline analysis establishing an empirical performance ceiling, a domain-adaptation (CORAL) baseline situating the proposed mechanism against feature-alignment approaches, a multi-seed ablation showing robustness to the prototype-weighting temperature, a full factorial ablation isolating the individual contribution of each architectural component, full 123-way identification with confusion-matrix analysis, and a floating-point-operation-based complexity analysis confirming suitability for resource-constrained edge deployment. Collectively, these results establish SNR-conditioned prototype computation as a mechanistically well-justified, computationally lightweight approach to channel-robust few-shot device authentication for dynamic IoT networks.