Description
Continuous physical-layer security monitoring is essential for practical quantum key distribution (QKD) systems, as device imperfections and attacks may compromise the fundamental assumptions underlying ideal security proofs. In recent years, machine-learning-based anomaly detection methods have demonstrated strong capabilities in multi-domain fault and attack localization. However, computationally large monolithic ensemble models introduce substantial inference overhead, making them difficult to deploy directly in latency-sensitive QKD post-processing environments. This research proposed a decoupled mixture-of-experts (MoE) routing framework for resource-efficient multi-domain attack detection in QKD. Specifically, the proposed method first standardizes photon-counting statistical features and projects them into a reduced 11-dimensional principal component subspace. A lightweight frontend classifier is then used to rapidly estimate the prediction confidence of each physical domain. When the confidence of a specific domain falls below a predefined threshold, the system activates only the corresponding domain expert, rather than sending the entire sample to a full-scale baseline model. The domain experts are designed for source, phase, detector, and attack anomalies, respectively. Through this physically decoupled routing mechanism, the framework avoids the multi-domain joint-probability penalty, in which local uncertainty in one physical domain forces the entire sample to undergo full-model evaluation. Simulation results based on a decoy-state BB84 QKD model show that the proposed method significantly reduces computational cost while maintaining high diagnostic accuracy. Under stringent confidence thresholds, the architecture achieves a macro-averaged F1-score of approximately 0.98 and reduces inference overhead by more than 50% compared with the monolithic baseline model. Explainable-AI analysis further shows that different domain experts attend to distinct principal component features, supporting the physical interpretability of the routing structure. These results indicate that physics-informed expert routing provides a scalable path toward continuous, interpretable, and low-latency anomaly detection in practical quantum communication systems, with strong potential for edge deployment.
| I am the presenting author | Yes |
|---|