Research On Continuous-Variable Quantum Key Distribution Defense Based On Machine Learning

As one of the most promising technologies in secure communication,continuous-variable quantum key distribution has theoretical unconditional security,but it is vulnerable to various attacks by hackers due to factors such as equipment defects.Traditional quantum attack detection requires professionals to judge whether an attack exists and what types of attacks exist through parameter estimation.Therefore,quantum attack detection based on machine learning can improve the universality of detection and realize automatic identification of attacks to avoid human consumption.This thesis focuses on the continuous-variable quantum key distribution system based on Gaussian modulation coherent state and performs high-precision intelligent detection of quantum attacks through machine learning.The main research contents are as follows:1.For the coexistence of multiple attacks,an actual continuous-variable quantum key distribution system for real-time intelligent monitoring experiment and simulation platform is built.And an attack detection module is added before the parameter estimation procedure of the system for automatic detection of attacks.In addition,the safety analysis of the improved real-time monitoring system is carried out.The simulation results show that the built monitoring system takes less than 0.05 ms in the actual attack identification and can fully adapt to the system of about 1-100 Mbps,realizing real-time intelligent detection of various known attack types and unknown attacks or threats.2.For the intelligent detection of various known attacks,a multiattack detection scheme based on neural networks is designed,which can identify different types of attacks autonomously and detect the coexistence of multiple attacks.According to the characteristics of multi-attacks having multiple output labels,a neural network model suitable for quantum multi-attack detection is designed,which can automatically identify and classify multi-attacks.The experimental simulation results show that the proposed BR-NN and LP-NN multi-attack detection models can perfectly predict known attacks,and the detection accuracy is 100%.Since the LPNN model considers the interaction of multiple attacks in the prediction stage,it is more suitable for detecting known attacks.Finally,a security analysis of the proposed attack detection scheme is carried out for the subsequent improvement of the model.3.For unknown threats and attacks,an anomaly detection module based on the One-class SVM algorithm is designed to detect unknown attacks.Further,the One-class SVM anomaly detection module,which has little impact on the model structure,is combined with the multi-attack detection model,thereby realizing high-precision perception of unknown attacks and unknown threats based on the detection of known attacks.The experimental simulation results show that in the improved multi-attack detection model,the detection accuracy of the BR-NN model is 99.86%and the detection accuracy of the LP-NN model is 98.7%,so the BR-NN model is used to detect unknown attacks.