Resisting Quantum Key Distribution Attacks Using Quantum Machine Learning

Abstract

The emergence of quantum computing poses significant risks to the security of modern communication networks as it breaks today's public-key cryptographic algorithms. Quantum key distribution (QKD) offers a promising solution by harnessing the principles of quantum mechanics to establish secure keys. However, practical QKD implementations remain vulnerable to hardware imperfections and advanced attacks such as photon number splitting and Trojan-Horse attacks. In this work, we investigate the potential of quantum machine learning (QML) to detect popular QKD attacks. In particular, we propose a hybrid quantum long short-term memory (QLSTM) model to improve the detection of common QKD attacks. By combining quantum-enhanced learning with classical deep learning, the model captures complex temporal patterns in QKD data, improving detection accuracy. To evaluate the proposed model, we introduce a realistic QKD dataset that simulates typical QKD operations, along with seven attack scenarios: Intercept-and-resend, photon-number splitting (PNS), Trojan-Horse attacks on the random number generator (RNG), detector blinding, wavelength-dependent Trojan-Horse and combined attacks. The dataset includes quantum security metrics such as quantum bit error rate (QBER), measurement entropy, signal and decoy loss rates and time-based metrics, ensuring an accurate representation of real-world conditions. Our results demonstrate the promising performance of the quantum machine learning approach compared to traditional classical machine learning models, highlighting the potential of hybrid techniques to enhance the security of future quantum communication networks. The proposed hybrid QLSTM model achieved an accuracy of 94.7.0% after 50 training epochs, outperforming classical deep learning models such as LSTM, CNN, ANN, random forest and RNN. However, our evaluation is conducted on a semi-realistic, simulation-generated decoy-state BB84 dataset with simplified models of attacks and device imperfections, so the reported performance should be interpreted as a proof-of-concept rather than a final assessment on field-deployed QKD systems.