“Have a look at the images in this paper on "adversarial examples": Explaining and Harnessing Adversarial Examples”
Explaining and Harnessing Adversarial Examples
对抗样本源于神经网络的线性,并给出快速生成方法
对抗样本是对数据集样本施加微小而针对最坏情形的扰动所得的输入,会让模型以高置信度输出错误答案。作者反对「非线性或过拟合」的早期解释,认为神经网络易受对抗扰动的主要原因是其线性性质,这既解释了对抗样本能跨架构和训练集泛化,也给出生成对抗样本的简单快速方法;用这些样本做对抗训练,可降低 maxout 网络在 MNIST 上的测试误差。
评论者引用它解释为何图像上不可察觉的微小改动就能让分类器出错,并强调这与过拟合无关、对抗样本能跨架构迁移,且可以通过对抗训练部分缓解。
学习路径
读这篇之前,按顺序读书架上的 2 篇 · 参考文献来自 OpenAlex,顺序和要点由 Claude Code 排出先读 Szegedy 等人首次报告对抗样本的论文,并了解 GoogLeNet 这个被攻击的模型,就能看懂 Goodfellow 为什么把原因归于线性并给出快速生成方法。
- 1
第一次发现微小扰动会让网络误分类,目标论文要解释的现象正是它提出的。
读对抗样本的发现和不同模型之间的可迁移性;L-BFGS 生成方式只需了解概念。
- 2
目标论文用 GoogLeNet 在 ImageNet 上演示扰动的效果,需要知道这个模型。
读Inception 模块的整体结构,不需要细看训练技巧。
- 3Explaining and Harnessing Adversarial Examples
目标论文:提出对抗样本源于高维线性,并给出 FGSM 与对抗训练。
读线性解释的论证、快速梯度符号法的公式,以及对抗训练作为正则化的效果。
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“A paper came out that explains this effect and a method of minimizing it: Explaining and Harnessing Adversarial Examples”
“Overfitting has nothing to do with it. See this paper: Explaining and Harnessing Adversarial Examples”
“Second some on here have posted about the Svegedy, Goodfellow, and Shlens paper Explaining and Harnessing Adversarial Examples which discusses the opposite effect.”
“Anyway this paper has a method for addressing this problem: Explaining and Harnessing Adversarial Examples”
“IIRC the follow up paper tested it or at least claimed it wasn't a solution.”
看全部 14 条 →
“He's also referring to this follow up paper which attempts to explain the issue and show that it exists in other models too: Explaining and Harnessing Adversarial Examples”
“Would this cost be analogous to sensitivity to adversarial examples in the case of convolutional neural networks [2]?”
“Anyway this issue has been explained and fixed somewhat in this paper: Explaining and Harnessing Adversarial Examples”
“There are other work in the literature describing faster algorithms to compute these perturbations, which makes it possible to use them while training. See, eg.: Explaining and Harnessing Adversarial Examples”
“Explaining and Harnessing Adversarial Examples is a Goodfellow et. al. (ie, the person who's Tweet you note above) paper where they show adversarial examples do apply across architectures”
“demonstrates this with logistic regression.”
“See for example Figure 1 in the seminal paper on adversarial attacks: Explaining and Harnessing Adversarial Examples”
“same as the old attacks adding noise to an image to have it misclassified despite clearly looking the same: Explaining and Harnessing Adversarial Examples (paper from 2014).”
- Explaining and Harnessing Adversarial Examples [problem with deep networks] ↗alexanderg · 2014-12-23 · 1 分
- Adversarial examples – Machine Learning [pdf] ↗psoto · 2015-04-10 · 2 分
- Explaining and Harnessing Adversarial Examples ↗pmalynin · 2015-12-16 · 3 分