Radio Frequency Fingerprint Identification Technology Study Based On Artificial Intelligence

In recent years,the widespread use of fifth-generation mobile communication technology in the Internet of Things has led to a dramatic increase in the number of terminal connections in communication networks,and communication security is urgently needed to ensure the quality of communication between terminals and to avoid data leakage.Traditional security technologies rely on high complexity keys,which can be broken by brute force.Therefore,based on the uniqueness and irreproducibility of the device's RF fingerprint,this thesis investigates RF fingerprint identification technology,the main content of which is as follows.This thesis investigates the mechanism of RF fingerprint formation.A transmitter model is constructed and the effects of the main modules such as frequency oscillator,quadrature modulator and power amplifier on the RF fingerprint and the characteristics of the overall RF fingerprint of the transmitter are analyzed.The theoretical analysis and numerical simulation results show that phase noise,amplitude imbalance in the in-phase and quadrature branches and non-linear effects are the main causes of the RF fingerprint.This thesis investigates RF fingerprint recognition techniques based on steady-state signals.A principle analysis of feature extraction techniques based on bispectral transform,time-frequency analysis and empirical modal decomposition is presented and an algorithmic flow is designed.The thesis first compares the performance of three sets of features of the same type,based on the generated signals simulated by the originating model,using typical classifiers for the selected bispectrum and the integral bispectrum,the short-time Fourier transform and the Wigner-Ville distribution and each empirical modal component.The theoretical analysis and numerical simulation demonstrate that the more typical and less redundant selected bispectrum,the Wigner-Ville distribution with better time-frequency aggregation,and the fourth and fifth groups with less noise and richer RF features Based on the preferred features,the paper also compares the performance of classical recognition algorithms such as support vector machines,Knearest neighbors network,bagged decision trees and convolutional neural networks,in which the selected bispectrum and empirical modal components have the highest recognition rate under support vector machines and the Wigner-Ville distribution has better performance under convolutional neural networks,resulting in alternative threeclass feature-classifier combinations.The thesis concludes with a performance validation and comparison of the three feature-classifier combinations based on real-world data from the software radio platform.Numerical simulations show that the selected bispectrum is more noise immune,the Wigner-Ville distribution has a higher recognition rate,and the empirical modal decomposition has the lowest time complexityThesis investigates RF fingerprint recognition techniques based on deep learning.For scenarios where the number of samples in the training set is insufficient,the training set is expanded based on the generative adversarial network,and the expanded recognition performance is tested using CNN.The results show that the expanded training set can effectively improve the recognition performance,but is limited by the feature richness of the original training set.For scenarios with different frequency bias and signal-to-noise ratio between the training and test sets,a domain adaptive network was constructed based on migration learning theory,and the recognition performance was compared with that of CNN using pre-compensation algorithm.The results show that both the domain adaptive network and the pre-compensation algorithm have certain migration capability,and the pre-compensation algorithm performs better when a priori knowledge is available.For the multi-category classification scenario,the classification results of multiple CNNs are combined using a voting mechanism based on one-verseone and one-verse-others integrated learning schemes,and the recognition performance is better than that of a single CNN on the measured signals.