Experimental robustness benchmarking of quantum neural networks on a superconducting quantum processor

Abstract

Quantum machine learning (QML) models, like their classical counterparts, are intrinsically vulnerable to adversarial attacks, hindering their secure deployment. Here, we report the first systematic experimental benchmark of robustness for 20-qubit quantum neural network (QNN) classifiers executed on a superconducting processor. Our benchmarking protocol features an efficient adversarial attack algorithm tailored for quantum hardware, enabling the diagnosis of QNN’s robustness across diverse datasets. The empirical upper bound extracted from our attack experiments deviates by only 3 × 10⁻³ from the analytical lower bound, providing strong experimental confirmation of our attack’s precision and the tightness of the fidelity-based robustness bounds. Furthermore, our quantitative analysis reveals that adversarial training mitigates sensitivity to targeted perturbations by regularizing input gradients, thereby significantly enhancing QNN robustness. Additionally, we observe that experimentally measured QNNs exhibit higher adversarial robustness than classical neural networks, an effect attributed to inherent quantum noise. Our work establishes the first scalable and experimentally accessible framework for robustness benchmarking, paving the way for secure and reliable QML applications.