Quantum Autoencoders for Anomaly Detection in Smart Grid Cybersecurity

Increased connectivity in cyber-physical smart grids enhances energy management, but it also expands the attack surface and makes security against sophisticated threats in highly imbalanced data environments increasingly challenging. Semi-supervised autoencoders for anomaly detection are trained only on normal operations and detect deviations from this baseline as anomalous, potentially malicious behaviour. In this work, we investigate quantum autoencoders (QAEs) for anomaly detection in smart grid transformer differential protection. We compare QAEs with various encoding strategies and architectures against conventional classical autoencoders (CAEs). Our results demonstrate that QAEs are competitive with CAEs while using fewer parameters. Select QAEs even show a higher recall, indicating a slightly better ability to detect false data injection attacks. Furthermore, gradient-based saliency analysis shows that quantum and classical autoencoders rely on similar features. Finally, to determine their trustworthiness for Noisy Intermediate-Scale Quantum (NISQ) hardware, we evaluate the robustness of the QAEs under a composite noise framework incorporating amplitude damping, phase damping, depolarization, and readout measurement errors. These results highlight the potential and provide a methodological framework for further investigation into using QAEs for anomaly detection in critical infrastructure.