Quantum Key Distribution (QKD) is a promising technique for ensuring long-term security in communication systems. Unlike conventional key exchange methods like RSA, which quantum computers could theoretically break [1], QKD offers enhanced security based on quantum mechanics [2]. Despite its maturity and commercial availability, QKD devices often have undisclosed implementations and are tamper-protected. This thesis addresses the practical imperfections of QKD systems, such as low and fluctuating Secret Key Rates (SKR) and unstable performance. By applying theoretical SKR derivations to measurement data from a QKD system in Poland, we gain insights into current system performance and develop machine learning (ML) models to predict system behavior. Our methodologies include creating a theoretical QKD model [2] and implementing ML models using tools like Keras (TensorFlow [3]). Key findings reveal that while theoretical models offer foundational insights, ML models provide superior accuracy in forecasting QKD system performance, adapting to environmental and operational parameters. This thesis highlights the limitations of theoretical models and underscores the practical relevance of ML models for QKD systems. Future research should focus on developing a comprehensive physical layer model capable of doing long-term forcasting of the SKR. Such a model could prevent an encryption system form running out of keys if the SKR drops significantly. In summary, this thesis establishes a foundational approach for using ML models to predict QKD system performance, paving the way for future advancements in SKR long-term predictions.

翻译：量子密钥分发（QKD）是一种有望确保通信系统长期安全性的技术。与RSA等传统密钥交换方法（理论上可被量子计算机破解[1]）不同，QKD基于量子力学原理提供增强的安全性[2]。尽管QKD技术已趋成熟并实现商业化，其设备的具体实现细节通常不予公开且具备防篡改保护。本论文针对QKD系统的实际缺陷展开研究，例如秘密密钥率（SKR）偏低且波动、性能不稳定等问题。通过将理论SKR推导应用于波兰某QKD系统的实测数据，我们深入分析了当前系统性能，并开发了机器学习（ML）模型以预测系统行为。我们的研究方法包括建立理论QKD模型[2]，以及使用Keras（TensorFlow[3]）等工具实现ML模型。关键研究结果表明：虽然理论模型提供了基础性见解，但ML模型在预测QKD系统性能方面具有更高的准确性，并能适应环境与运行参数的变化。本论文揭示了理论模型的局限性，并强调了ML模型对于QKD系统的实际应用价值。未来研究应致力于开发能够进行SKR长期预测的综合性物理层模型。此类模型可在SKR显著下降时防止加密系统耗尽密钥。综上所述，本论文建立了利用ML模型预测QKD系统性能的基础性方法，为未来实现SKR长期预测的技术进步铺平了道路。