Few-Shot Cross-Domain Radio Frequency Fingerprint Identification: A Comparative Study of Deep Learning Architectures with Prototypical Networks
Abstract
Radio Frequency Fingerprint Identification (RFFI) uses the hardware impairments created during manufacturing such as power amplifier non-linearity, IQ imbalance, and oscil-lator phase noise - as a device-specific fingerprint to support non-cryptographic, passive wireless authentication. This paper compares the performance of five deep learning architectures: A baseline CNN, an improved CNN, a CNN-LSTM hybrid, ResNet and Domain-Adversarial Neural Network (DANN) on the WiSig dataset which consists of 130 COTS WiFi transmitters from four time-invariant domains. Every model is tested on five signal representations (time-domain magnitude, FFT, IFFT, STFT spectrogram, and IQ decomposition), with three-stages of protocol: within-domain accuracy, few-shot fine-tuned adaptation (on prototypical networks) and zero-shot cross-domain transfer. When prototypical network episodic fine-tuning is added to ResNet, the best results are obtained with 97.4% within domain accuracy and 95.9% fine-tuned cross domain accuracy after 3 shot adaptation, compared to the 87-88% accuracy of standard classifiers. The label-efficient prototypical adaptation is found to be robust through all the architectures with accuracy above 92% after fine-tuning. The results prove the usefulness of ResNet-based prototypical networks for cross domain RFFI and suggest a stepping stone toward scalable, adaptive wireless device authentication systems.