End-to-end Communication System Based On Autoencoder And Its Robustness Research

With the rapid development of wireless communication technology,intelligent communication has gradually become the mainstream trend.The end-to-end communication system uses the deep learning method to realize the joint optimization of the wireless physical layer communication system,which provides new possibilities for the design and optimization of the wireless communication system.At present,the end-to-end communication system has shown great application potential in the wireless physical layer.This paper focuses on the end-to-end communication system based on the autoencoder model.The model is not based on any classical coding and detection methods,but regards the wireless physical layer as a whole,that is,the entire end-to-end communication system is constructed as a deep neural network model for information reconstruction and joint optimization.In order to deeply study the robustness of the model,this paper designs an improved C&W attack algorithm based on end-toend communication systems,and proposes corresponding defense measures against the perturbations generated by the algorithm.Firstly,the basic knowledge of deep learning and its application status in the wireless physical layer are briefly summarized,and the end-to-end communication system structure based on the autoencoder model is mainly studied.The BLER performance of the system under different network structures is simulated and analyzed,which lays the foundation for the research in subsequent chapters.Then,the attack algorithms of the end-to-end communication system are mainly studied,and the universal adversarial attacks based on the end-to-end communication system is deeply researched and simulated.The attack effect of the algorithm is verified from the perspective of white-box and black-box through experimental simulation.Aiming at the problem that the existing algorithm needs to repeatedly calculate the distance value,an improved C&W attack algorithm based on end-to-end communication system is proposed.Simulation results show that the proposed adversarial attack algorithm can achieve effective attacks on end-to-end communication systems under both white-box and black-box attacks.Finally,the defense measures of the end-to-end communication system are mainly studied,and the development status of classical defense measures in the field of computer vision is summarized.The defense measures against C&W black-box attacks are studied from three perspectives.The C&Wbased adversarial example training method achieves the defense effect by adding a certain proportion of adversarial perturbations to the system channel during the training phase.The simulation results show that the adversarial example training method with a scale factor of 0.4 can effectively improve the robustness of the end-to-end communication system without affecting the BLER performance of the original system as much as possible.But this approach incurs a performance penalty when the end-to-end system is not attacked.Thus,the defense methods based on additional random noise and an improved loss function are proposed.The simulation results show that the proposed methods can defend against C&W black-box attacks,and the improved loss function defense method can also improve the BLER performance of the system itself to a certain extent.Aiming at the problem that the above algorithms cannot completely eliminate the impact of attack algorithms on system performance,a discrimination-based perturbation rectifying defense method is further proposed,which almost completely eliminates the negative impact of C&W black-box attacks on system performance without affecting the BLER performance of the original system when it is not attacked.