Neural network for excess noise estimation in continuous-variable quantum key distribution under composable finite-size security

Parameter estimation is a critical step in continuous-variable quantum key distribution (CV-QKD), as the statistical uncertainty from a finite data size leads to pessimistic worst-case bounds that drastically reduce the secret key rate and range. While machine learning techniques have been proposed for this task, they have lacked the rigorous statistical framework necessary for integration into a composable security proof. In this work, we bridge this gap by introducing a statistically rigorous framework for using neural networks for parameter estimation in CV-QKD with quantifiable composable security. We develop a neural network estimator for the excess noise and, crucially, derive its worst-case confidence interval using a delta method approach, ensuring the estimation fails with a probability not exceeding εPE. This allows the network to be integrated into a parameter estimation protocol that is operationally equivalent to the standard maximum likelihood method but yields significantly tighter parameter bounds. Our numerical results demonstrate that this method provides substantially more precise estimates, which directly translates into a higher secret key rate and extended transmission distance over a fiber channel under a collective Gaussian attack. This work establishes that machine learning can be securely and effectively harnessed to overcome a key performance limitation in practical CV-QKD systems.