Research On Radio Frequency Fingerprint Extraction And Identification

With the continuous popularization of wireless communication technology and the vigorous development of the Internet of Things technology,wireless communication has played an irreplaceable role in both military and civilian aspects.It has become an indispensable part of modern society.However,due to its broadcast nature,wireless networks are more vulnerable to large-scale malicious attacks than traditional wired networks,and their security issues cannot be ignored.Radio Frequency Fingerprint(RFF)technology analyzes the communication signals of wireless devices to extract the RFF of the device,thereby identifying and authenticating different transmitters.RFF represents the transmitter's hardware differences due to differences in circuit design and manufacturing tolerances,which can be parasitic in communication signals.Different devices have different RF fingerprints.RFF extraction and identification technology can provide higher security performance for civilian wireless networks.They can also offer many new identification systems and methods based on existing research on Specific Emitter Identification(SEI)and communication station identification.Therefore,the research on RFF extraction and identification technology has profound theoretical significance and great practical value in both military and civilian fields.This thesis has carried out some researches on several key issues in the field of RFF extraction and identification in wireless communication,including RFF modeling and simulation,blind frequency offset estimation based on Differential Constellation Trace Figure(DCTF),Multi-Sampling Convolutional Neural Network(MSCNN)and Denoising Auto Encoder(DAE)methods combined with deep learning technology,as well as semi-steady signal definition,detection and cause analysis.The innovative achievements of this thesis include:1.Researched on time-domain RFF modeling and simulation.The modeling of the RFF generation mechanism is the basis of subsequent research and guides how to select fingerprint characteristics.According to the proposed general digital communication system framework,the influence of each component in the transmitter on RFF is analyzed,and a corresponding RFF model is established.The tolerances of four important parameters in the relevant communication standards in the field of RFF are summarized.Based on the Root Mean Square(RMS)Error Vector Magnitude(EVM)tolerance,the upper and lower bounds of the DAC DC offset,I/Q gain imbalance,I/Q quadrature offset,I/Q filter offset,oscillator phase noise,and non-linearity of the power amplifier for two typical modulation methods,QPSK and 16-QAM,are theoretically derived.The changes of the constellation diagram under various critical bounds are simulated,and the influences of single and mixed RFF parameters are analyzed,which provides reasonable parameter guidance for future modeling and simulation in the RFF field.2.Proposed a blind frequency offset estimation algorithm for MPSK signals based on DCTF.The constellation of the baseband signal with the residual frequency offset is constantly rotated,but its corresponding DTCF has a fixed deflection.The quasi-symmetry of the DCTF generated by the MPSK signal without shaping filter in the AWGN channel is theoretically proved.It is found that the axis of symmetry passes through the origin and the rotation angle with respect to the in-phase axis is only related to the frequency offset and the differential interval.It is proposed to estimate the rotation angle of the axis of symmetry by minimizing the asymmetry energy term,thereby estimating the frequency offset.Simulation results show that the method performs well at medium to high SNRs on AWGN channels and flat Rice channels,and its Normalized Mean Square Error(NMSE)is slightly higher than the modified Cramer-Rao bound.Compared with the traditional algorithms,the frequency offset estimation range is greatly improved,and the estimation range becomes dependent on the receiver sampling rate.Finally,the method is successfully adapted for frequency offset estimation of ZigBee devices.3.Proposed an MSCNN RFF Identification method based on the adaptive Region-Of-Interest(ROI)selection.In order to extract the frequency offset-independent features,this thesis proposed a fine frequency offset estimation algorithm based on signal despreading to eliminate the residual frequency offset after the coarse frequency offset estimation.Aiming at the unstable semi-stable phenomenon caused by ZigBee device sleep mode switching,an SNR-adaptive ROI selection algorithm was proposed to effectively trade-off the effective classification information of the signal and the semi-stable region jitter.MSCNN framework is proposed to improve the traditional single-sampling-rate neural network method,which includes a downsampling phase,a local convolution phase,and a fully connected phase.It can automatically and efficiently extract multi-scale features of the same size.It significantly improves the identification rate of the device.In the line-of-sight scene with an SNR = 30 dB,the identification accuracy of 54 CC2530 devices is 97%,and in the non-line-of-sight scene with an SNR = 15 dB,the identification accuracy is 84.6%4.Studied the semi-steady state behavior and its causes.Starting from the definition of transient and steady-state,the traditional transient signal,steady-state signal,and newly discovered semi-steady state signal are redefined.The characteristics of non-data transient signals and semi-stable signals in the data segments for the ZigBee devices with and without power amplifiers are compared and analyzed in detail.Based on the difference in the distribution of three kinds of differential signals in the semi-stable region and the steady-state region,a Bayesian ramp-up change-point detection method based on the differential phase sliding window standard deviation is proposed to detect the semi-steady endpoint.Experimental results show that the method is very stable on most devices and extremely robust to noise.Finally,the effects of power amplifiers,balun networks,Low-Dropout voltage(LDO)regulators,and transmitting powers on semi-steady signals are analyzed through various experimental studies.Finally,the causes of the semi-steady state are explained.5.Proposed a general RFF framework based on Denoising Auto Encoder,and a partial stacking strategy for ZigBee devices.The traditional SNR enhancing method of multi-frame stacking is improved,and a novel method of stacking the same symbol in a single frame to improve the SNR is proposed.Taking into account the RFF difference between the semi-steady-state signal and the steady-state signal of ZigBee devices,only the steady-state region is stacking by symbols and then concatenated to the semi-steady-state region,which can effectively combine the RFF identification advantages of these two regions at different SNRs.In view of the poor performance of traditional deep learning RFF methods at low SNRs,an improved method based on a Denoising Auto Encoder is proposed,which is trained by minimizing the weighted reconstruction errors and classification losses at the same time.The experimental results show that in the low SNR scenarios(-10 dB ?5dB)under the AWGN channel,compared with the traditional CNN method,the classification accuracies of 27 ZigBee devices are improved by 14% ?23.5%.