Abstract: Continuous-variable quantum key distribution (CVQKD) has theoretical unconditional security, which is based on the assumption that the physical devices at the sender and receiver operate perfectly and are secure and reliable. However, in practical CVQKD systems, the eavesdropper can launch attacks from three aspects—source, channel, and detection end—by exploiting the physical flaws inherentin actual devices, thereby compromising the practical security of the system. Although corresponding defense strategies have been developed for some practical quantum attacks, each strategy can only defend against specific attack types, lacking a universal defense approach effective against most attacks. By combining machine learning techniques with attacks detection in the CVQKD, we propose a practical attacks detection scheme in this work based on the K-nearest neighbors (KNN) algorithm. This scheme extracts features from the optical pulses of the CVQKD system, trains a KNN prediction model through learning, and ultimately deploys the model at the receiver end of the CVQKD system to detect practical quantum attacks. Simulation results demonstrate that the proposed attack detection scheme can effectively identify various typical quantum attacks targeting CVQKD, with both precision and recall rates exceeding 98%.

Key words: quantum information, continuous-variable quantum key distribution, practical security, attacks detection, machine learning