University Archives

Radio frequency (RF) fingerprinting identifies wireless transmitters by leveraging unique hardware-induced imperfections present in transmitted signals. Unlike traditional identification methods that rely on higher-layer identifiers, RF fingerprinting operates at the physical layer by analyzing subtle variations in in-phase and quadrature (I/Q) signal characteristics. In recent years, machine learning (ML) and deep learning have advanced RF fingerprinting performance by enabling models to learn representations directly from the raw signal data. However, the effectiveness and robustness of different ML architectures remains an open question, particularly under varying signal-to-noise ratio (SNR) conditions and edge-deployment environments.

This systematic literature review examined ML approaches used for RF fingerprinting, focusing on convolutional neural networks (CNNs), hybrid CNN/recurrent neural network (RNN) models, and attention-based Transformer architectures. Studies using commonly referenced datasets (including ORACLE, DARPA RFMLS, WiSIG) were analyzed to evaluate how architectural design, benchmarking practices, and evaluation metrics influence reported performance. The review covered key trends in model development, identified limitations in current methodologies, and examined how models performed across different SNR conditions. The findings suggest that while CNN-based architectures remain strong baselines, emerging architectures such as Transformers show potential for improved representation learning.