ROBUSTNESS OF DEEP LEARNING-BASED RADIO FREQUENCY FINGERPRINTING UNDER JAMMING AND NOISE CONDITIONS

Abstract

This thesis investigates the robustness of deep learning–based Radio Frequency (RF) fingerprinting systems under adversarial interference conditions. RF fingerprinting identifies wireless devices using hardware-induced signal variations, but these features can be affected by interference and environmental distortions. Building on the FingerJam framework, this study experimentally evaluates the impact of controlled friendly jamming and receiver-side Additive White Gaussian Noise (AWGN) on RF fingerprint classification. Experiments were conducted using USRP X310 Software Defined Radios and a ResNet-18 convolutional neural network to classify signals from five transmitters operating at 900 MHz with Binary Phase Shift Keying (BPSK) modulation. The baseline experiment shows that classification accuracy decreases progressively as jamming power increases, dropping from near-perfect performance on clean signals to approximately random-chance accuracy (20%) at high interference levels. Additional experiments demonstrate that lower temporal resolution (10³ samples per image) improves robustness due to larger training dataset sizes, while a mini-batch size of 64 provides the best classification performance. A comparison with receiver-side AWGN reveals a sharper degradation pattern, with classifier performance collapsing when the signal-to-noise ratio falls below approximately 21 dB. These results highlight the vulnerability of RF fingerprinting systems to interference and demonstrate that friendly jamming can effectively disrupt RF fingerprint extraction while preserving communication functionality.

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