“To say it's both fascinating and interesting really understates it. Really neat stuff!”
Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images
人眼无法辨认的噪声图,却能让深度网络以 99.99% 的置信度分类
作者用进化算法和梯度上升,针对在 ImageNet 或 MNIST 上训练的卷积网络生成人眼完全无法识别的图像,网络却以接近确定的置信度(最高 99.99%)将其归入某个类别,称之为「愚弄图像」。这说明当时的 DNN 与人类视觉存在重要差异,并对 DNN 视觉的泛化性提出疑问。
HN 上的人把它当作神经网络图像识别容易被欺骗的经典例证,用来说明模型与人类视觉的差距,或讨论对抗人脸识别和生物识别的可能性。
学习路径
读这篇之前,按顺序读书架上的 5 篇 · 参考文献来自 OpenAlex,顺序和要点由 Claude Code 排出先了解 ImageNet 分类器与卷积网络可视化,再读对抗样本的发现,最后看如何生成人眼无法辨认却被高置信分类的图像。
- 1
目标论文攻击的是在 ImageNet 上训练的、接近人类水平的分类器,需要了解这个基准。
读任务设定和人机对比部分。
- 2
- 3
通过梯度生成类别图像,与目标论文用优化方式生成欺骗图像的思路相近。
读类别得分对输入的梯度上升生成图像的做法。
- 4
- 5
- 6Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images
目标论文:用进化算法和梯度方法生成人眼无法辨认、网络却高置信分类的图像。
读两种生成方法(进化算法、梯度上升)与高置信度结果的实验设置及其启示。
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“For some background, see Deep Neural Networks are Easily Fooled which has some cool visuals of pictures that make neural networks to strange things.”
“I still don't get the hype around”
“He showed some neat results from these models (some published in Deep Neural Networks are Easily Fooled, and some from [LINK]).”
“Witness, for example, all the recent "we can fool deep learning image recognition systems" papers.”
“For example, folks have already figured out how to trick deep neural nets:”
看全部 11 条 →
“want to add a link to this paper which shows how far we still have to go”
“This Deep Neural Networks are Easily Fooled could be one way.”
“or some drawbacks like the ones explained by Deep Neural Networks are Easily Fooled or [LINK]”
“I'm reminded of a paper for which the authors generated different pictures of static that fooled neural network image classifiers into confidently identifying them as different objects: Deep Neural Networks are Easily Fooled”
“Here is an article explaining this problem: Deep Neural Networks are Easily Fooled”
- Deep Neural Networks Are Easily Fooled ↗sinwave · 2014-12-09 · 183 分
- Deep Neural Networks Are Easily Fooled [pdf] ↗cgtyoder · 2014-12-11 · 1 分