A Hybrid Quantum-Classical Neural Network for the Detection of Quantum Hacking Attacks in CVQKD
Abstract
Flaws in Quantum Key Distribution are inherent due to the imperfect physical devices on which it is implemented, these flaws leave an attack surface for hackers to exploit as seen in the LOIA exploits and calibration attacks.
In an effort to minimize this attack surface, this study proposes a novel HQCNN architecture for identifying various quantum hacking attacks by using a classical CNN layer, extracting and reducing key statistical features from signals, combined with a VQC for precise recognition and robust decision making of complex attack patterns. After conducting five-class classification experiments simulating a CVQKD system environment, we achieve a 6% increase in accuracy over the classical ResNet model while significantly reducing the number of parameters used in the model. This demonstrates a dual improvement in model compactness and detection accuracy.
This novel implementation effectively provides a high level of security, making it virtually impossible to reverse engineer the original signal distribution or the key generation parameters even in complex adversarial environnement.