Beyond universal attack detection for continuous-variable quantum key distribution via deep learning

Removing quantum hacking attacks is a highly challenging task for quantum key distribution (QKD) since one needs to correctly detect the types of attacks. There is a big gap between the idealized theoretical model and the practical physical system, which results in various hacking strategies exploited from security loopholes. To bridge this gap, previous continuous-variable (CV) QKD systems adopted a universal statistical model to represent different dimensional characters of the physical layer. The challenge is to discriminate different attacks from the exposed key physical details. Here we present a semisupervised deep-learning method to detect the known attacks or potential threats that are hard to be captured in previous CVQKD systems. The proof-of-principle experiment result shows that it could break the present limitations that different attacks are relatively independent, and key features computed from previous threshold-based methods are not correlated with each other. We anticipate that our proposal will give insight into how quantum hacking attacks can be detected under the real system modeling constraints, and will enable the present quantum communication systems to reach a higher level of practical security without adding new assumptions.