Radio Frequency Fingerprinting: Deep Learning-based Approaches

Abstract

In this dissertation, we focus on the authentication issues that arise when numerous Radio Frequency (RF) devices are wirelessly connected in an Internet of Things (IoT) environment. Modern cryptographic key-based Media Access Control (MAC) layer authentication systems pose significant security threats if keys are lost and require complex algorithms that are not suitable for IoT environments. To address these issues, RF Fingerprinting technology, which authenticates devices using unique RF signal characteristics at the physical layer, has gained attention. RF Fingerprinting starts from the premise that it is significantly more costly to replicate the unique characteristics of RF signals, thereby playing a crucial role in enhancing the security of wireless networks. This dissertation presents two significant research findings related to RF Fingerprinting and discusses the challenges and future research directions for achieving system integrity. The first study proposes an RF Fingerprinting method capable of identifying signal transmitters in a frequency-hopping spread spectrum (FHSS) network, one of the best existing RF security systems. To achieve this, Signal Fingerprints (SF) of the hopping signals were extracted and transformed into spectrograms representing the time-frequency behavior of the SFs. These spectrograms were then trained using a designed Deep Inception Network (DIN)-based signal classifier. As a result, the system was able to identify signal transmitters with 97% accuracy for hopping signals, and the algorithm for detecting network attackers showed an Area Under the Receiver Operating Characteristic (AUROC) curve performance of 0.99. The second study emphasizes the importance of public key management in IoT environments and introduces an RF-based Public Key Generator (RF-PubKG) model. The proposed model, by projecting RF feature clusters into cryptographic sequences, achieved 97.2% accuracy at a 20dB Signal-to-Noise Ratio (SNR), which further improves to 99.6% in noiseless conditions. Furthermore, low-correlation analysis of the generated public key sets confirmed the reliability and independence of the learning-based public key system, and a proof of concept of the RF-PubKG-based Rivest–Shamir–Adleman (RSA) digital signature system demonstrated that it could effectively operate as Public Key Cryptography (PKC) without the need for Public Key Infrastructure (PKI). These results are expected to simplify public key management in IoT environments and significantly improve the efficiency of the digital signature verification process. Finally, we discuss the challenges and future research directions for the integrity of the RF Fingerprinting system. The contribution of this study spans from existing methods operating with analog feature keys in the real domain to the RF-PubKG based on public keys operating in the finite field. Furthermore, the research direction includes the development of sensor intelligence systems based on digitized analog feature keys operating in the digital domain. This involves discussing the challenges of public operation in the current system and proposing future research directions to overcome these challenges.