Abstract:
Quantum key distribution (QKD) supports secret key exchange to enable data exchange with guaranteed security but remains vulnerable to key exchange interruption caused by physical-layer threats, such as high-power jamming attacks. In particular, in a fiber-based QKD network equipped with optical switching capabilities, a high-power jamming signal injected into a single link can be optically switched and propagated to other links, resulting in the interruption of key exchange on multiple links. To address this challenge, we introduce a novel metric, the maximum number of affected requests (maxNAR), which quantifies the worst-case impact of a single physical-layer attack. Based on this, we investigate a new problem: Routing and Wavelength Assignment with Minimal Attack Radius (RWA-MAR). The objective of RWA-MAR is to assign routes and wavelengths to QKD requests in such a way that the maximum number of requests disrupted by any single physical-layer attack is minimized. We first formulate the problem using an integer linear programming (ILP) model to minimize maxNAR. Due to computational limitations of ILP, we develop a scalable deep reinforcement learning (DRL) solution that dynamically balances security (minimizing maxNAR) and resource efficiency. The DRL solution is designed to utilize key caches (defined as QKP) to minimize the maxNAR, as the QKP can save the keys for future use. Extensive simulations across diverse topologies and workloads show that our approach achieves a significant reduction in maxNAR (over 50%) compared to a baseline that performs RWA without considering the reduction of an attack impact. Meanwhile, our approach maintains reasonable resource efficiency in key delivery, marking a step toward robust and attack-aware QKD network design.