[论文] Defending Quantum Classifiers against Adversarial Perturbations throug...

## 论文概要 **研究领域**: ML/Quantum **作者**: Sagnik Chakraborty, Malay Singh, Arpit Jain **发布时间**: 2026-04-30 **arXiv**: [2604.28176](https://arxiv.org/abs/2604.28176) ## 中文摘要 机器学习模型可以从数据样本中学习以高效执行各种任务。当数据样本被对抗性地操纵时，例如插入精心设计的噪声，可能导致模型出错。量子机器学习模型同样容易受到此类对抗攻击，特别是在使用变分量子分类器进行图像分类时。虽然存在一些对抗扰动的有前景的防御方法，如使用对抗样本训练，但它们面临实际限制。例如，它们不适用于无法使用对抗样本训练或可能导致模型对某种攻击过拟合的场景。本文提出一个无需对抗训练的防御框架，利用量子自编码器通过重建来净化对抗样本。此外，我们的防御框架提供一个置信度指标来识别可能无法被量子自编码器净化的对抗样本。大量评估表明，我们的防御框架在对抗攻击下的预测准确率显著优于最先进方法（提升高达68%）。 ## 原文摘要 Machine learning models can learn from data samples to carry out various tasks efficiently. When data samples are adversarially manipulated, such as by insertion of carefully crafted noise, it can cause the model to make mistakes. Quantum machine learning models are also vulnerable to such adversarial attacks, especially in image classification using variational quantum classifiers. While there are promising defenses against these adversarial perturbations, such as training with adversarial samp... --- *自动采集于 2026-05-02* #论文 #arXiv #ML/Quantum #小凯