Research On Radiation Source Identification Of Communication Individuals Based On Deep Learning

As wireless networks are increasingly deployed autonomously,the identification of transmitters is becoming increasingly important.There are various forms of attacks on wireless networks,and impersonation attacks are one of the most important and most threatening.In this attack,an illegal transmitter can copy most of the identification information,such as passwords and Media Access Control(MAC)addresses to spoof the device,which puts wireless security at serious risk,in which legitimate security credentials are obtained by an adversary,thus compromising security.The existence of this threat underscores the need for technologies that identify and authenticate the identity of transmitters.However,identifying and characterizing transmitters in real time is still a challenging task,and the common approach today is to use security mechanisms such as cryptographic keys to authenticate the transmitter identity.In this paper,the principles and feature extraction and deep learning algorithms for the identification of individual radiation sources for communication are studied in depth,with the following main research elements.1.This paper studies the mechanism of radio frequency fingerprint formation of wireless communication transmitters and builds a real-world data set collection system.And built the Gnu Radio-Hack RF one-3900 A integrated real-world transmitter data acquisition platform,and formed the real-world data set for RF fingerprint identification tasks,laying the foundation for subsequent research.2.In this paper,a DWT-ReliefF-based feature extraction algorithm is proposed.Since the recognition accuracy of traditional RF fingerprint recognition methods is not high,this paper firstly starts from the feature extraction stage and proposes a ReliefF feature selection algorithm based on Discrete Wavelet Transform(DWT).The algorithm firstly performs wavelet decomposition on the wireless signal,obtains the wavelet energy spectrum features of the signal,constructs the wavelet energy spectrum features,then uses ReliefF algorithm to analyze the weight of the wavelet energy spectrum,eliminates the features below the threshold weight,and uses the remaining wavelet energy values to construct the feature subset,and the experimental results prove that the DWT-ReliefF feature extraction algorithm has higher accuracy than Hilbert yellow transform,power spectrum density feature extraction method has a higher recognition accuracy of 86.68%.3.A sample expansion method based on Generative Adversarial Network(GAN)is proposed based on the problem of small number of samples in the measured dataset.The method uses the unique adversarial learning method of generative adversarial network(GAN)to train the generator and obtain the generated data with similar features to the real sample feature space.The effect of the feature space of the generated data with different sample expansion ratios on the recognition accuracy is analyzed in conjunction with the experimental results,and the problem that the limited improvement in recognition accuracy of the expanded dataset is influenced by the feature richness of the original data is also discussed.4.In this paper,an RF fingerprint recognition method based on improved Deep Residual Shrinkage Network(DRSN)is used.Since the RF fingerprint recognition task is prone to low recognition accuracy,this paper uses an improved network deep residual systolic network model of deep residual network(Res Net),which incorporates a special soft thresholding module in the residual learning unit,aiming to extract useful features and eliminate redundant features from the received noise-laden signal.The model has 90.16% recognition accuracy on the measured dataset of this study,which is 3.32% better than the traditional deep residual network recognition accuracy,and also has higher recognition accuracy compared with CNN,VGG,LSTM and other network structures.This paper also compares the recognition accuracy on the public dataset CIFAR-10,and the Deep Residual Shrinkage Network(DRSN)used in this paper achieves an accuracy of 90.07%,which is higher than the Deep Residual Network...