RF Fingerprint Identification Using Power Amplifier Nonlinearity as a Feature
Sankari, Juhani (2025)
2025
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-07-31
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202507317973
https://urn.fi/URN:NBN:fi:tuni-202507317973
Tiivistelmä
Radio frequency fingerprint identification (RFFI) is a technique used to identify wireless transmitters based on subtle, hardware-induced variations in their emitted signals. These imperfections, which are unique to each device, serve as fingerprints that can be used for tasks such as identification, authentication, and intrusion detection.
Power amplifier (PA) nonlinearity–based radio frequency (RF) fingerprints are a promising candidate for RFFI tasks, as the distortion introduced by the PA is inherently device-specific, stable over time, and difficult to replicate or spoof. These nonlinear characteristics arise from physical imperfections and manufacturing variability, making them well-suited for identifying individual transmitters.
This thesis focuses on the role of PA nonlinearity as a reliable source of RF fingerprints. Using a simulation-based approach, a population of transmitters is modelled, each characterised by a distinct PA response derived from real-world measurements. The influence of nonlinearity is systematically controlled using a backoff parameter, allowing exploration of its impact on RF fingerprints. An additive white Gaussian noise (AWGN) channel with adjustable noise levels is used for data augmentation. Classification is performed using various convolutional neural network (CNN) architectures with spectrogram-based feature representations. The objective is to examine how PA nonlinearity, spectrogram type, noise level in training data, and model architecture influence classification performance. Results are evaluated in terms of device-level and manufacturer-level accuracy.
The results demonstrate that PA nonlinearity is a promising discriminative feature for RF fingerprinting. It proves effective in both traditional direct classification models and in hybrid approaches combining embedding models with a separate classifier. In the best-performing scenario, the direct method achieved 85.5% accuracy at the device level and 100% accuracy at the manufacturer level when classifying 20 devices from two different manufacturers.
Power amplifier (PA) nonlinearity–based radio frequency (RF) fingerprints are a promising candidate for RFFI tasks, as the distortion introduced by the PA is inherently device-specific, stable over time, and difficult to replicate or spoof. These nonlinear characteristics arise from physical imperfections and manufacturing variability, making them well-suited for identifying individual transmitters.
This thesis focuses on the role of PA nonlinearity as a reliable source of RF fingerprints. Using a simulation-based approach, a population of transmitters is modelled, each characterised by a distinct PA response derived from real-world measurements. The influence of nonlinearity is systematically controlled using a backoff parameter, allowing exploration of its impact on RF fingerprints. An additive white Gaussian noise (AWGN) channel with adjustable noise levels is used for data augmentation. Classification is performed using various convolutional neural network (CNN) architectures with spectrogram-based feature representations. The objective is to examine how PA nonlinearity, spectrogram type, noise level in training data, and model architecture influence classification performance. Results are evaluated in terms of device-level and manufacturer-level accuracy.
The results demonstrate that PA nonlinearity is a promising discriminative feature for RF fingerprinting. It proves effective in both traditional direct classification models and in hybrid approaches combining embedding models with a separate classifier. In the best-performing scenario, the direct method achieved 85.5% accuracy at the device level and 100% accuracy at the manufacturer level when classifying 20 devices from two different manufacturers.