2026/07/27 by Ruixue Yang, Lin Bi, Gopal Verma
paper · doi:10.1088/1402-4896/ae9143
Abstract In practical continuous-variable quantum key distribution (CVQKD) systems, the co-fiber transmission of classical and quantum signals introduces complex noise and crosstalk, leading to inaccurate parameter estimation, modulation mismatch, and key rate degradation, threatening security. Existing methods struggle to determine whether the model output conforms to physical laws and security requirements. This paper proposes a Deep Semantic Model (DSM) that integrates a Physics-Informed Neural Network (PINN) and a Transformer, embedding channel physics into the learning process to achieve adaptive modulation variance adjustment under noise. A key innovation is the introduction of a "semantic credibility" metric to measure whether the model output conforms to the physical mechanisms and security boundaries of CVQKD, ensuring the adjustment process remains secure. The adaptive strategy under the GG02 protocol achieves an absolute gain of approximately 0.082bits/pulse , representing a relative improvement of 21%. This verifies that the proposed theory effectively improves the practicality and robustness of CVQKD, providing reliable support for its practical application.