Deep Reinforcement Learning-Driven Optimization of End-to-End Key Provision in QKD Systems

The advent of quantum computing poses significant threats to traditional cryptographic methods, necessitating the development of secure communication techniques such as quantum key distribution (QKD). Despite advancements in QKD, including enhanced fiber optic technologies and multi-key distribution systems, substantial challenges persist in the efficient management of cryptographic keys within QKD networks. Existing heuristic approaches, such as greedy algorithms, often fail to address the complex requirements of key allocation, particularly for provisioning end-to-end keys essential for secure communication between distant nodes. This paper introduces a reinforcement learning (RL)-based method for end-to-end key provisioning in QKD networks. The proposed approach dynamically optimizes key allocation using the state and usage patterns of the network. Specifically, the RL framework integrates graph attention networks and long short-term memory networks to model intricate relationships and temporal dependencies within the network. This integration enables a more efficient and adaptive key distribution. Comparative analyses demonstrate that the RL-based method significantly improves session key availability and allocation efficiency. It outperforms traditional greedy algorithms by minimizing session interruptions and reducing unused quantum keys. These results provide valuable information on the practical implementation of RL-based key provisioning strategies in real-world QKD applications.