PUF-Assisted Radio Frequency Fingerprinting Exploiting Power Amplifier Active Load-Pulling

• Published in: IEEE transactions on information forensics and security 2024, Vol.19, p.5015-5029
• Main Authors: Li, Yuepei, Xu, Kai, Zhang, Junqing, Gu, Chongyan, Ding, Yuan, Goussetis, George, Podilchak, Symon K.
• Format: Article
• Language: English
• Subjects: Active load-pulling; Algorithms; Classification; Communication system security; convolution neural network (CNN); Fingerprinting; Hardware; Integrated circuit modeling; Internet of Things; Machine learning; non-linear memory effect; Nonlinearity; physical unclonable function (PUF); Power amplifiers; Radio frequency; radio frequency fingerprinting (RFF); Signal to noise ratio; Wireless sensor networks
• ISSN: 1556-6013; 1556-6021
• DOI: 10.1109/TIFS.2024.3389570

This paper presents a novel radio frequency fingerprint (RFF) enhancement strategy by exploiting the physical unclonable function (PUF) to tune the RF hardware impairments in a unique and secure manner, which is exemplified by taking power amplifiers (PAs) in RF chains as an example. This is achieved by intentionally and slightly tuning the PA non-linearity characteristics using the active load-pulling technique. The motivation driving the proposed research is to enlarge the RFF feature differences among wireless devices of same vendor, in order to massively improve their RFF classification accuracy in low to medium signal to noise ratio (SNR) channel conditions. PUF is employed to dynamically tune the PA's RFF feature which guarantees the security since the PUF response cannot be cloned. Specifically, a ring oscillator (RO)-based PUF is implemented to control the PA non-linearity by selecting unique but random configuration parameters. This approach is proposed to amplify the distinctions across same model PAs, thereby enhancing the RFF classification performance. In the meantime, our innovative strategy of PUF-assisted RFF does not noticeably compromise communication link performance which is experimentally tested. The resulting RFF features can be extracted from the received distorted constellation diagrams with the help of image recognition-based machine learning classification algorithms. Extensive experimental evaluations are carried out using both cable-connected and over-the-air (OTA) measurements. Our proposed approach, when classifying eight PAs from a same vendor, achieves 11% to 24% average classification accuracy improvement by enlarging the RFF feature differences arising from the PA non-linearity.