Research On HRRP Target Recognition And Adversarial Attacks Based On Deep Neural Networks

With the rapid development of information technology,artificial intelligence has become the core driving force for the new military revolution and plays an increasingly important role in the field of national defense.Integrating artificial intelligence technology with radar automatic target recognition(RATR)technology to enhance the detection and early warning capabilities of incoming targets is of great significance to improve the situational awareness on the battlefield.Furthermore,as the high-resolution range profile(HRRP)represents the projection of the complex returned echoes from the target scatting centers onto the radar line-of-sight(LOS),contains the target structure signatures,and has the advantages of easy acquisition,storage and processing,the HRRP-based target recognition has received intensive attention from the radar automatic target recognition(RATR)community.By considering the engineering background of the radar HRRP recognition,this dissertation studies the radar HRRP feature extraction(target recognition)and adversarial attacks based on deep neural networks.The main research work is summarized as follows:1.In order to explore the correlation between range cells and extract the structured discriminative features in HRRP,a novel approach to address HRRP RATR and rejection problem based on convolutional neural network(CNN)is proposed.In the proposed model,we add a reconstruction network to the basic CNN recognition network framework to achieve the target rejection problem.Through the combination of the recognition and reconstruction networks,we integrate the target recognition with outlier rejection task together.Furthermore,since the spectrogram representation of HRRP is more informative than time domain HRRP,besides using one-dimensional CNN to handle time domain HRRP,we also design a twodimensional CNN model for the HRRP spectrogram representation.Experimental results on measured HRRP data show that the proposed CNN model outperforms many state-of-art methods for both recognition and rejection tasks.2.The influence of HRRP time-frequency representation on the recognition results is studied,and a CNN with attention mechanism is proposed for HRRP target recognition tasks.Detailedly,we first employ two time-frequency transform methods,i.e.,short-time Fourier transform and continuous wavelet transform,to analyze the HRRP samples and compare their recognition performance in CNN networks.On this basis,to solve the problem of parameter selection when using HRRP time-frequency representation,a radar HRRP target recognition method based on attentional CNN with multi time-frequency representation is proposed.The attention mechanism in this method automatically merges the features extracted by CNN under different HRRP time-frequency representations,which not only solves the time-frequency parameter selection issue,but also helps improve the recognition rates.3.In order to make the features extracted from the deep network not only effective,but also resistant to time-shift sensitivity,a CNN and bidirectional recurrent neural network(Bi RNN)mixture model,called CNN-Bi RNN,is proposed for the HRRP target recognition task.In the proposed method,the CNN is utilized to explore the spatial correlation of HRRP data and extract expressive features followed by a Bi RNN taking the full consideration of temporal dependence between range cells.Furthermore,in order to enhance the robustness to misalignment,an attentional mechanism is also introduced in CNN-Bi RNN to allow the model focus on the discriminative target area during feature extraction.The proposed CNN-Bi RNN model combines the advantages of CNN and RNN.In other words,compared with the CNN model,CNN-Bi RNN also considers the temporal dependence in HRRP,which can alleviate the HRRP time-shift problem,while compared with the RNN model,CNN-Bi RNN introduces CNN for feature extraction,which can effectively improve the recognition rate.Experimental results on measured HRRP data demonstrate the effectiveness and the robustness to misalignment of the proposed method.4.The adversarial attack problem in deep network based HRRP recognition is studied,and a robust digital adversarial attack method is proposed.Deep networks are easily fooled by adversarial examples,and the research on the adversarial examples can help us to fool the deep networks based target recognition system.In this study,we first introduce several classic digital adversarial sample methods,and compare their attack performance in the HRRP recognition networks.Then,in order to improve the robustness and practicability of the adversarial attack,we propose a robust digital adversarial attack method,which generates a general,local adversarial perturbation for network attacks through optimization.Compared with the classic digital attack methods,the adversarial perturbation generated by the proposed method is more robust and practical,and provides a basis for future study on physical attacks on the recognition system.