Research On Key Technologies Of Physical Layer Identification Of Wireless Devices

Wireless communication and network technologies have been widely applied in critical areas such as social production,transportation,aerospace,military,and national defense.The deployment and application of emerging networks,such as the new generation of wireless communication and the Internet of Things(Io T)networks,further expand the application scope of wireless networks and promotes access to massive wireless terminals.Wireless networks are more vulnerable to information eavesdropping,tampering,and forgery attacks than wired networks because of their wireless medium’s inherent openness,broadcast,and signal superimposition characteristics.Identity-based attacks such as spoofing attacks,impersonation attacks,and Sybil attacks can lead to the disclosure of sensitive information on wireless networks and even threaten network infrastructure security.The physical layer authentication constructs and verifies the wireless device’s unique identity through its physical layer’s inherent characteristics,thus ensuring legitimate access and communication.It enhances and supplements the traditional upper layer-based wireless network security scheme.As the key to physical layer authentication,the research on physical layer identification of wireless devices has recently mainly focused on introducing artificial intelligence technologies such as machine learning to identify wireless devices.This type of research is primarily divided into two categories: schemes using shallow classifiers based on physical layer features extracted by feature engineering and schemes using deep learning based on raw signal samples.However,for the identification scheme of shallow classifiers,there is currently a lack of performance-limit research on the physical layer features on which they depend.In addition,the existing physical layer feature types cannot meet the need to improve identification accuracy,and expanding the physical layer feature types is urgent.Physical layer identification for deep learning lacks solutions that can cope with signal quality and channel changes.Based on the above discussion,this dissertation studies the performance of the device identification based on the physical layer feature of carrier frequency offset(CFO)and shallow classifiers.The device identification based on the physical layer feature of normalized Shannon entropy of the horizontal visibility graph(HVGE)and its combination with other features is studied.Time-frequency fusion physical layer identification based on deep learning is studied.The main contribution and innovations of this dissertation are as follows:(1)The identification accuracy,the number of identifiable devices,and the corresponding supremum in different feature ranges,signal-to-noise ratios,and channel environments are studied for the identification scheme using the physical layer feature of CFO.Using the Max-Min Distance Analysis(MMDA)criterion,the identification scheme’s analytic expression of the accuracy supremum was analyzed and derived.The supremum of the number of identifiable devices under the constraints of a given feature range,signal-to-noise ratio(SNR),and accuracy is analyzed.The supremum of the identification accuray of the multi antennas communication system is also studied.The simulation platform of WLAN communication is established,and the theoretical analysis of supremum is verified.The results show that the feature range and feature estimation method are the main factors affecting the ability to distinguish features,and multi antennas can improve the supremum of the identification accuracy.(2)To improve the accuracy of the identification scheme based on physical layer features,the HVGE is used as the physical layer feature of the direct sequence spread spectrum(DSSS)communication system,and the multi-feature combination identification is studied under different Gaussian white noise and multipath channels.A method for extracting the physical layer feature,i.e.,HVGE,of the preambles of the wireless devices is proposed.Signal preprocessing methods such as sampling truncation and double-downsampling are proposed to double the efficiency of feature extraction.The software-defined radio is used to build a physical layer identification experimental platform to collect the preamble frame data of 50 wireless LAN devices in different channel environments.The correlation between HVGE and CFO,amplitude error,IQ imbalance,phase error,and synchronization error is analyzed through experimental analysis and comparison.The experiments show that the Pearson correlation coefficient between HVGE and the above five features is less than 0.28.The identification accuracy of different feature combinations under different SNRs and multipath channels was studied.The identification accuracy of the combination of five features under the Gaussian channel was increased by 11% after adding HVGE and 5% in multipath channel environments.Finally,the model size,training time and inference time of different identification models are analyzed and compared.(3)A time-frequency fusion identification method is proposed to integrate the raw IQ samples and power spectral density information to reduce the influence of signal quality and channel change on the accuracy of physical layer identification based on deep learning.Input fusion,feature fusion,and decision fusion identification schemes are proposed based on a fully connected deep neural network,convolutional neural network,and recurrent neural network.The experimental platform was developed using software-defined radio technologies.We evaluated the identification schemes’ accuracy and their model’s training and inference time overhead in different channel environments.The results show that the identification accuracy of the proposed fusion scheme is higher than that of the existing scheme under different SNRs of the Gaussian channel and multipath channel.The identification accuracy of the fusion scheme at low SNR is 40% higher than that of the scheme using only power spectral density and 5% higher than that of the raw IQ samples-based scheme at high SNR.The feature fusion scheme can maintain the highest identification accuracy in the multipath channel environment.Compared with the scheme that only uses raw IQ samples,the fusion scheme can reduce the model’s training time by 30% while ensuring the frame-by-frame identification inference time requirements.This dissertation focuses on the physical layer identification of wireless devices based on machine learning and systematically studies the performance supremum of the identification method based on CFO,which provides a reference for improving the identification performance.The proposed physical layer feature HVGE and their multi-feature combination identification can effectively improve the accuracy of device identification based on physical layer features and shallow classifiers.The proposed time-frequency fusion physical layer identification scheme based on deep learning can obtain the highest identification accuracy under different channels,thereby reducing the influence of the identification scheme on signal quality.The collected frame preamble data sets of wireless devices in three indoor channel environments can provide basic data support for research in fields such as physical layer authentication and identification and enrich existing public data sets.This dissertation can provide theoretical,and application references for physical layer identification and has a prominent role in promoting the research of device authentication and identification in the field of physical layer security.