Routing Optimization of QKD-Networks using Machine-Learning Based Prediction

Abstract:
Ensuring the security of digital infrastruc-ture is an essential field of work in today’s communications research. Nowadays secure communication is typically based on the computational complexity of algo-rithms, which may become vulnerable in the future due to emerging new technologies like quantum computers. Quantum key distribution (QKD) is a promising long-term approach to secure data transmission especially for critical infrastructure like communications of authorities. QKD can ensure information-theoretical security based on the fundamental laws of quantum physics. For sufficient reach also in long-haul networks the concept of “trusted nodes” is considered as a practical variant in this contribution. Before (symmetric) encryption of the data packets can start, first secure keys need to be exchanged between two QKD-devices, which currently is only possible at a very limited key rate. These keys are stored in the keystore of each QKD-device. A lack of available keys in a secure node consequently will lead to a denial of a request, which needs to be avoided. In this contribution we present traffic prediction-based routing optimization applied to a meshed long-haul QKD network topology. The traffic prediction is imple-mented using machine learning with a Long Short-Term Memory (LSTM) model. For the prediction, we train the LSTM model to forecast the future traffic matrices. Based on the prediction result the weights for the routing algorithm can be adapted dynamically. If the predicted traffic load is above the remaining capacity of a keystore, weights of the edges are adapted in advance. As a result, the optimal route is adapted before a keystore runs empty, which otherwise would lead to a blocking of the request. In this way the keystore proactively has the possibility to regenerate. To evaluate the performance of the LSTM based optimization, we compare it with a simple hop-count based shortest path algorithm and with an optimization algorithm based on a (hypothtical) perfect prediction.