Research On Offensive And Defensive Of Continuous-Variable Quantum Key Distribution System Based On Machine Learning

In the practical continuous-variable quantum key distribution system,the gap between the real-world device and the theoretical assumptions leads to security loopholes in the system.The practical security of the system is threatened by multiple attacks.Therefore,analyzing the potential vulnerabilities of the system,modeling specific attacks,and seeking relevant defense measures are the main problems that the quantum key distribution protocol faces in its application.During the communication,the change in ambient temperature,channel transmission loss,and eavesdropping will cause quantum signal disturbances.Thus,it is difficult to determine what kind of attack the system is subject to by directly measuring the quantum signal.Machine learning is adept at handling multidimensional and multivariate data and is robust in dynamic environments.Therefore,machine learning can be used to avoid the effects of undesirable realistic conditions to some extent and achieve the detection of attacks.The thesis initially explores the application of machine learning to the offensive and defensive of the continuous-variable quantum key distribution systems from different perspectives,and the main contents are as follows.(1)A calibration attack detection scheme based on the Gradient Boosting Decision Tree for the practical continuous-variable quantum key distribution system is proposed.In the practical system,Eve can launch a calibration attack to steal the secret key through the local oscillator.The existing defenses need improvement in terms of increased system complexity or the number of keys to be sacrificed.In the thesis,we model calibration attacks.Then,based on the impact of calibration attacks on system transmission parameters,we introduce a Gradient Boosting Decision Tree model to distinguish normal communication data from calibration attack data without increasing the system complexity.Some statistics describing samples are introduced,such as mean,variance,kurtosis,median absolute deviation,etc.We make a dataset based on these statistics and use it to train the model.In the meantime,the key role they play in attack detection is demonstrated and analyzed.The results show that the scheme sacrifices fewer secret keys while achieving 99.2% accuracy without increasing the system complexity.(2)A sifting scheme based on Generative Adversarial Networks for the practical continuous-variable quantum key distribution system is proposed.In the sifting step,the legitimate parties discard the apparently useless portion of the raw data to form a sifted key.In practical systems,useless data consists of abnormal data and the key measured by the legitimate party on different bases.The thesis proposes a sifting scheme that can preliminarily sift abnormal keys in the sifting step.Firstly,the features of parameters are extracted to make a dataset based on the differences between the system transmission parameters in abnormal and normal communication.Then,the sifting model is obtained to monitor the abnormal data by training the model which is based on Long Short Term Memory networks and uses a variant of Generative Adversarial Networks,Ganomaly as the basic architecture.The results show that the model has a high performance in distinguishing normal communication from local oscillator intensity attacks,calibration attacks,saturation attacks,and denial of service attacks while requiring fewer raw keys.Our strategy uses short samples for monitoring and achieves high accuracy and high speed.The measure helps the practical systems preliminarily sift the useless raw data.By employing our measure,the secret key rate of the model can be improved when under abnormal communication.