Single-step phase-compensation scheme for continuous-variable quantum key distribution based on self-supervised recurrent neural networks
Phys. Rev. Applied 25, 044087 – Published 29 April, 2026
DOI: https://doi.org/10.1103/b1fl-khyn
Abstract
Continuous-variable quantum key distribution (CV-QKD) is highly attractive for secure optical communication due to its compatibility with commercial infrastructure. In local local oscillator systems, time-division multiplexing is widely employed to compensate for phase drift caused by independent lasers. However, conventional approaches rely on a two-step compensation process that sacrifices a portion of the raw key data, thereby directly reducing the final secret key rate. To address this, we propose a single-step phase-compensation scheme using a self-supervised recurrent neural network based on the Noise2Void framework. By leveraging the temporal correlation of phase drift against the independence of measurement noise, our “blind-spot” network effectively denoises pilot phase observations without requiring clean ground truth labels. Experimental validation over an 80-km fiber link demonstrates that the proposed method reduces excess noise to 0.0035 shot-noise unit (SNU), significantly lower than the 0.0183 SNU for the traditional two-step scheme. By achieving “zero raw key overhead,” it boosts the secret key rate to 225 kbps under the asymptotic regime and 75 kbps under the finite-size regime, corresponding to increases of 46% and 140% compared with the conventional scheme. This work provides a robust, cost-effective solution for enhancing the performance of practical CV-QKD systems.
