| With the rapid development of wireless sensor network technology,the information security of the Internet of Things(Io T)has received more and more attention.Criminals use hackers to illegally invade the Io T system,which may lead to the leakage of private information or sensitive data,or even cause the system to crash,thus causing huge economic losses to users.The authentication technology based on the radio frequency fingerprint is a scheme to ensure the security of the wireless system at the physical layer.The radio frequency fingerprint is an authentication identification generated by the inherent impairments of the device hardware.It does not depend on any protocol or mathematical method as well as additional information.It is unique and cannot be tampered with.In addition,this kind of scheme also has the characteristics of low cost and low power consumption,and is suitable for various wireless devices.Most of previous researches on radio frequency fingerprinting only focus on the classification of known devices without the identification of unknown devices.In reality,due to the sudden and accidental intrusion of rogue devices,we cannot obtain samples of rogue devices in advance.Therefore,the identification of rogue devices is more challenging and more realistic.This thesis proposes a radio frequency fingerprint authentication scheme based on adaptive filter coefficients,which involves both the classification of known devices and the identification of unknown devices.The coefficients of the adaptive filter is determined by the difference between the actual signal and the reference signal,so it can reflect the subtle differences between the signals of different devices.At the same time,the adaptive filter is a general module in the communication system,so no additional modules need to be designed and imported for fingerprints extraction.Zig Bee devices widely used in Io T systems are used as experiment devices.Signals from 54 devices are preprocessed end adaptive filter coefficients are extracted from them as radio frequency fingerprints.Various algorithms are selected for signals with different SNR.The performances of these algorithms are analyzed by comparing the accuracy and running time of them.The main work of this thesis is as follows:1.Preprocess the collected original Zig Bee signal to make it as same as the reference signal as much as possible.If the input signal of the adaptive filter differs too much from the reference signal,the filter cannot converge.These preprocessing include coarse frequency offset compensation,synchronization,fine frequency offset compensation,resampling,and phase fitting.2.Adjust the parameters of adaptive filters such as least mean square(LMS),recursive least square(RLS),affine projection(AP),and block LMS(BLMS)filters,and observe the error convergence results.Experimental results can prove that signal errors such as direct current offset,I/Q relative amplitude gain,and I/Q relative phase offset will affect the adaptive filter coefficients,thereby verifying the reasonableness of the adaptive filter coefficients as radio frequency fingerprints.3.For the task of radio frequency fingerprint classification,three supervised algorithms are compared: support vector machine(SVM),random forest(RF),and k-nearest neighbor algorithm(KNN),and combined the classification accuracy and algorithm training speed under different signal-to-noise ratios(SNR).4.For the task of radio frequency fingerprint identification,three unsupervised algorithms are compared: one-class support vector machine(OCSVM),isolated forest(IF),and local outlier factor algorithm(LOF),and analyze the identification accuracy of legal devices,the identification accuracy of rogue devices and the training speed under different SNR,and give the optimal solution.Experiments show that OCSVM is the best choice in terms of training speed and accuracy. |
