“Note that adversarial examples generalize across models.”
Practical Black-Box Attacks against Machine Learning
只看输出标签、训练替代模型,就能攻击远程托管的深度网络
论文首次演示攻击者在不知道模型内部结构和训练数据的情况下,控制远程托管的 DNN:攻击者只能观察目标模型对所选输入给出的标签,用这些标签训练本地替代模型,再用替代模型构造对抗样本。对 MetaMind 托管的 DNN,84.24% 的对抗样本被误分类;用逻辑回归替代模型攻击 Amazon 和 Google 的模型,误分类率分别为 96.19% 和 88.94%,并能规避此前认为会增加攻击难度的防御策略。
HN 上的人引用它说明攻击者不需要访问目标网络,只要拿输入输出训练一个替代模型,就能构造对抗样本。
学习路径
读这篇之前,按顺序读书架上的 3 篇 · 参考文献来自 OpenAlex,顺序和要点由 Claude Code 排出先理解对抗样本与物理世界中的攻击,再看对模型的黑盒窃取,最后读只凭输出标签训练替代模型的黑盒攻击。
- 1
- 2
- 3
- 4Practical Black-Box Attacks against Machine Learning
目标论文:训练替代模型,以黑盒方式攻击远程网络
读合成数据扩充训练替代模型,以及对抗样本从替代模型迁移到目标模型
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“Interestingly, you can treat the NN as a black box (ie, not look at individual weights or even the architecture) and still derive adversarial cases:”
“No, this paper by Papernot et al shows how to do blackbox attacks without knowledge of model internals.”
“You don't need access to the network, you can just build another one.”
“create and train a local model as a stand-in for the actual model using the inputs and outputs of the actual model.”
“Nicholas Papernot [1] et al. has shown how this can be an attack on an unknown network provided you have access to inference results produced by the target network [2].”
- Practical Attacks against Deep Learning Systems using Adversarial Examples ↗Houshalter · 2016-02-23 · 98 分
- Practical Black-Box Attacks Against Machine Learning ↗signa11 · 2017-08-17 · 1 分
- Practical Black-Box Attacks Against Machine Learning ↗imheretolearn · 2020-06-21 · 2 分