Why RF Fingerprinting Needs Better Data, Not Bigger Models

Recent advances in Radio Frequency Fingerprint Identification have demonstrated impressive classification performance, often relying on deep neural networks. However, such approaches remain too computationally intensive for deployment on resource-constrained IoT devices. In this work, we revisit Fully Connected Neural Networks (FCNNs) and show that a simple one-hidden-layer FCNN can achieve competitive performance, requiring at least 5× fewer FLOPs and inference time than state-of-the-art architectures, with only a marginal drop in F1-score (0.02% to 0.21% depending on the dataset, for 16 transmitters or fewer). More importantly, our results reveal that despite their simplicity, our FCNNs exhibit comparable resilience to channel and device variability comparable to more complex architectures. This suggests that the current emphasis on increasing model size may be misguided. Rather than continuously increasing model depth or sophistication, our results highlight the importance of reexamining how datasets are designed and used in evaluation.