基于变分量子分类器的量子对抗攻击生成算法

Quantum Adversarial Attack Generation Algorithm Based on Variational Quantum Classifiers

  • 摘要: 量子分类器在扰动攻击下的脆弱性是量子机器学习中的基本理论问题之一。量子分类器的脆弱性是指其随着量子系统规模增大而更容易因为一些微小的扰动而分类错误的性质。这种微小扰动也被称为量子对抗攻击,而如何生成尽可能小的扰动使得量子分类器失效仍是一个开放问题。针对这一问题,提出了一种新的量子对抗攻击生成算法——量子混淆算法。该算法利用量子分类器关于输入数据的梯度信息来生成扰动,从而使得已训练好的量子分类器失效。数值仿真结果表明,与已有的量子对抗攻击方法相比,量子混淆算法可以通过更小的扰动实现对抗攻击,为理解分类器的有效性和脆弱性提供了新的思路。

     

    Abstract: The vulnerability of quantum classifiers under adversarial attacks is one of the fundamental problems in quantum machine learning. The vulnerability of quantum classifiers refers to the property that a quantum classifier may be failed by small perturbations when the quantum system scales up. Such perturbations are also known as quantum adversarial attacks. How to generate small perturbations to fail a quantum classifier is still an open problem. To address this problem, we present a new quantum adversarial attack generation method, the quantum confounding algorithm, which generates perturbations that fail the trained quantum classifier by utilizing the gradient information of the quantum classifier with respect to the input data. Numerical results demonstrate that, compared with the existing quantum adversarial attack generation methods, our quantum confounding algorithm can generate significantly smaller perturbations that lead the quantum classifier to malfunction. This provides a new perspective in understanding the effectiveness and the vulnerability of quantum classifiers.

     

/

返回文章
返回