Impact of Mobility in Radio Frequency Fingerprinting

Abstract

Radio Frequency (RF) fingerprinting is a technique that utilizes the hardware-caused imperfections in wireless transmitters to facilitate device identification at the physical layer. Unlike traditional identification methods that utilize higher-layer credentials, RF fingerprinting utilizes the minute analog properties hidden inside the transmitted signals, providing an extra layer of protection against spoofing attacks and device impersonation. Recent breakthroughs in deep learning have made it possible to automatically learn distinctive features from raw in-phase and quadrature (I/Q) signal samples without relying heavily on manual design of signal features. Nevertheless, the learned RF fingerprint robustness in practical wireless environments, especially when the transmitter is mobile, is still an open issue. This thesis proposes the use of autoencoders for transmitter identification based on raw I/Q signal chunks, with special emphasis on the effect of mobility on fingerprint robustness. A reconstruction-based unsupervised learning approach is adopted, where an autoencoder is trained on signals from a single transmitter and the reconstruction error is used as the identification criterion. Comprehensive experiments are performed on multiple transmitters in both static and mobile environments. The effect of chunk size, training time, and combinations of transmitters on the distributions of reconstruction errors is systematically analyzed. The experimental results show that the reconstruction error is low for the training transmitter and higher for other transmitters, making it possible to distinguish between them. Increasing the chunk size enhances distinguishability, and increasing the number of training epochs has little effect after convergence. When mobility is considered, channel fluctuations may degrade the distinguishability of fingerprints, especially when mobile data is considered during training. To overcome this shortcoming, an image representation of I/Q chunks is presented. The experimental results validate the effectiveness of the proposed image representation in improving distinguishability among transmitters in mobility scenarios. This work presents a thorough experimental study of reconstruction-based RF fingerprinting and emphasizes the importance of signal representation and training parameters for reliable transmitter identification in a wireless environment

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Science and Engineering