Deep Learning Based Transmitter Identification using Power Amplifier Nonlinearity

Abstract

The imperfections in the RF frontend of different transmitters can be used to distinguish them. This process is called transmitter identification using RF fingerprints. The nonlinearity in the power amplifier of the RF frontend is a significant cause of the discrepancy in RF fingerprints, which enables transmitter identification. In this work, we use deep learning to identify different transmitters using their nonlinear characteristics. By developing a nonlinear model generator based on extensive measurements, we were able to extend the evaluation of transmitter identification to include a larger number of transmitters beyond what exists in the literature. We were also able to study the impact of transmitter variability on identification accuracy. Additionally, many other factors were considered including modulation type, length of data used for identification, and type of data being transmitted whether identical or random under a realistic channel model. Simulation results were compared with experiments which confirmed similar trends.

Transmitter Classification with Supervised Deep Learning

“…They are setup so they function mostly in their linear region but, even there, their response curve is not perfectly linear. The parameters corresponding to this are slightly different from one radio chip to another, even amongst the same product line [3].…”

“…They develop an architecture called a multi layer perceptron made of a network of small neural networks and introduce the use of wavelet transforms with the goal of learning to classify with as few training samples as possible. The work in [3] focuses on amplifier characteristics and measures non linearity variations between 7 USRP cards. The data is used to train a network on hundreds of simulated devices.…”

Bayesian Neural Networks for Identification and Classification of Radio Frequency Transmitters Using Power Amplifiers’ Nonlinearity Signatures

“…(iii) Power amplifier distortions: Power amplifier (PA) nonlinearities are generally modeled using AM/AM (amplitude to amplitude) and AM/PM (amplitude to phase) curves. The work in [15] investigates the accuracy of several deep neural network architectures for the identification of RF fingerprints resulting from AM-AM conversion with frequency-domain representations of captured data. To generate similar data from different transmitters, the Volterraseries-based statistical model is implemented to synthesize PA nonlinearity to train the machine learning model for transmitter identification [15], [16].…”

“…The work in [15] investigates the accuracy of several deep neural network architectures for the identification of RF fingerprints resulting from AM-AM conversion with frequency-domain representations of captured data. To generate similar data from different transmitters, the Volterraseries-based statistical model is implemented to synthesize PA nonlinearity to train the machine learning model for transmitter identification [15], [16]. In addition to inherent hardware features, external perturbations coupled into the digital domain for data transmission have been proposed [17] for transmitter identification.…”

Radio Frequency Fingerprint Identification Based on Multi-Intervals Differential Constellation Trace Figures

“…But when faced with a large number of devices, such as large-scale wireless sensor networks, machine learning methods usually cannot achieve a high identification accuracy. As a very effective end-to-end learning method, deep learning has also been widely used in the RFF field recently [11,12,13]. Convolutional neural networks (CNN) is translation invariant and can learn spatial hierarchies of patterns [14,15].…”