RF Fingerprint Extraction Based On Convolutional Neural Network And FPGA Accelerator Design

Due to the widespread adoption of personal mobile devices and the rapid advancements in the industrial Internet of Things,wireless communication technology is gaining significant significance in both military and civilian spheres.Unlike wired communication channels,wireless connections are more susceptible to external malicious attacks due to their inherent openness,thus making ensuring communication security an increasingly critical concern for the industry.With the popularity of personal mobile terminals and the rapid development of industrial Internet of Things technology,wireless communication technology is playing an increasingly important role in military and civilian fields.Compared with wired communication,the openness of wireless communication leads to it being more vulnerable to malicious attacks from outside,and how to ensure the security of communication is getting more and more attention from the industry.Radio frequency fingerprint technology extracts the differences in transmit signals caused by hardware circuit design and manufacturing tolerances as the RF fingerprint of the transmitting device,and thus realizes the authentication identification between transmitters.As a physical feature of communication devices,RF fingerprint has stability,uniqueness and long-time invariance,and has emerged as a fresh area of interest for researchers in the field of communication security.Deep learning has powerful feature extraction capability,and RF fingerprint technology based on deep learning can automatically extract RF fingerprints while judging the recognition results,but it has the problems of low recognition accuracy and poor robustness.Therefore,this thesis investigates and analyzes a large number of deep learning algorithms in the field of RF fingerprinting,proposes a deep metric learningbased RF fingerprint extraction algorithm for multipath scenarios combined with KNN classification method,and deploys it on FPGA to achieve forward computation acceleration of RF fingerprint extraction network.The main research contents are as follows:Firstly,this thesis introduces the mechanism of RF fingerprint generation and gives its basic features and recognition model.This thesis provides a brief overview of the fundamental components of artificial neural networks,forward and backward propagation algorithms.Additionally,the residual convolutional neural network utilized in this thesis is thoroughly explained to facilitate comprehension of its fundamental architecture and operational principles.Secondly,an RF fingerprint extraction algorithm for multipath scenarios is proposed to address the problem that multipath channels can cause RF fingerprint corruption under current deep learning extraction algorithms.The algorithm architecture proposed relies on the prevailing Orthogonal Frequency Division Multiplexing(OFDM)technology,firstly,the OFDM transmission model and de-channelization are introduced to minimize the signal distortion resulting from multipath interference,and then the OFDM transmit and receive links are built in a laboratory multipath scenario using Universal Software Radio Peripheral(USRP).Then,the network model for RF fingerprint extraction and the classification method for recognition are introduced.Finally,after the preceding processing,the algorithm’s recognition results and performance are evaluated using actual data sets.Then,to deploy the above network on FPGA and achieve hardware acceleration,this thesis uses batch normalized layer fusion and 8-bit quantization to optimize the original network model and analyze the parallelism of the computational process,and finally obtains the hardware architecture design of the computational and data scheduling paths suitable for this network.Finally,to improve the problem of slow computation of RF fingerprint extraction network,the above network is deployed on Field Programmable Gate Array for hardware acceleration.The original network model is first optimized by batch normalized layer fusion and 8-bit quantization,and then the computational process of the computationally intensive convolutional layer is analyzed for parallelism to obtain the overall hardware architecture and key unit design suitable for the RF fingerprint extraction network,and finally the specific performance index is obtained through off-board experiments and compared with other FPGA gas pedals in the literature to compare and analyze their performance advantages and disadvantages.