按年被提到
20232023 · 12 次2026
视觉、语音与图像生成›扩散模型·arXiv:2301.13188 · cs.CR

Extracting Training Data from Diffusion Models

从扩散模型中提取训练数据
Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramèr, Borja Balle, Daphne Ippolito, Eric Wallace · 2023

扩散模型会记住并复现训练图像,且可被提取

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 17 人首次 2023-01-31 · 最近 2025-07-19
17不同的人
8活跃月份
16讨论串
294被引(OpenAlex)

论文证明扩散模型会记住训练集中的个别图像,并在生成时原样输出。作者用「生成再过滤」的流程,从最先进的模型中提取出一千多个训练样本,包括真人照片和注册商标 logo;还训练了数百个不同设置的扩散模型,分析建模与数据选择如何影响隐私。总体结论是扩散模型的隐私性远不如 GAN 等早期生成模型,缓解这些漏洞可能需要隐私保护训练方面的新进展。

HN 为什么提它

评论者在争论 AI 是否会记忆、复制训练数据以及版权问题时,把它当作扩散模型会复现训练图像的实证依据,也有人质疑其用刻意构造的提示来提取。

实证研究难度 3 / 5约 40 分钟经得起时间
读之前最好知道
1扩散模型(如 Stable Diffusion)
2成员推断与隐私攻击
3CLIP 嵌入
4记忆与泛化

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
推荐babel_ · 2023-02

“I think the paper is well worth the read, it's not particularly long (much is references and appendices), and nicely written”

推荐一读,论文不长,写得好,主要内容是提取训练数据
Stable Diffusion found to emit some images from training at generation time原评论 ↗
推荐pksebben · 2023-02

“For those following along, (1) is the best paper I could find talking about extracting images from SD.”

称这是自己找到的讲从 Stable Diffusion 提取图像的最好论文
Stable Attribution原评论 ↗
批评Ukv · 2024-01

“for papers like this, where they're generating hundreds of millions of images and digging into CLIP embeddings to construct prompts that intentionally recreate images”

OpenAI says it's "impossible" to create AI models without copyrighted material原评论 ↗
引用wokwokwok · 2023-02

“In this work, we show that diffusion models memorize individual images from their training data and emit them at generation time.”

引它的结论:扩散模型会记住训练图像,并在生成时输出
An IP Attorney’s Reading of the Stable Diffusion Class Action Lawsuit原评论 ↗
引用nimithryn · 2023-02

“IANAL either, but there are attacks you can use to extract the original training data from the models: Extracting Training Data from Diffusion Models.”

指出有攻击方法可以从模型中提取原始训练数据
Getty Images v. Stability AI – Complaint原评论 ↗
引用aix1 · 2023-03

“We already know that, at least for some image models, the answer to that question is”

说明至少对某些图像模型,训练数据已被证明能被重建
Launch HN: Flower (YC W23) – Train AI models on distributed or sensitive data原评论 ↗
看全部 15 条 →
引用pwinnski · 2023-04

“My mistake! I mis-remembered the original Google paper[0], which stated only that the image was in the Stable Diffusion training set.”

更正自己误记,Google 这篇论文只说图像在 Stable Diffusion 训练集中
Stable Diffusion copyright lawsuits could be a legal earthquake for AI原评论 ↗
引用vitehozonage · 2023-04

“often to the point where training data can be extracted [1][2]”

引它说明扩散模型常到能提取出训练数据的程度
Record Labels Panic as A.I. Generated Song Becomes Hit Single原评论 ↗
引用kmeisthax · 2023-04

“There's a paper from Google and Princeton about regurgitation happening in Stable Diffusion and Imagen: Extracting Training Data from Diffusion Models”

指出 Google 和普林斯顿的论文研究了 Stable Diffusion 与 Imagen 的训练图像复现
GitHub Copilot emits GPL code原评论 ↗
引用raincole · 2023-11

“It's already proven that you can reconstruct at least a small fraction of the training set from diffusion models.”

称已证明至少能从扩散模型重建一小部分训练集
Artists Lose First Round of AI Art Copyright Infringement Case原评论 ↗
引用Marazan · 2023-11

“Look at the Ann Graham Lotz image and tell me that isn't the source image being reproduced in a lossy manner.”

让人看 Ann Graham Lotz 的图,称那就是有损地复现源图
Let's Not Flip Sides on IP Maximalism Because of AI原评论 ↗
引用johnnyanmac · 2024-01

“some exploits discovered lately seem to suggest that this isn't just "reading a page" and then moving on: Extracting Training Data from Diffusion Models”

引它说明扩散模型并非只是“读一眼”就走,可能记住并存储原始素材
OpenAI says it's "impossible" to create AI models without copyrighted material原评论 ↗
引用snakeyjake · 2024-01

“Here's one-- Extracting Training Data from Diffusion Models Extracting Training Data from Diffusion Models”

举它为例,说明 AI 生成器会泄露原始训练图像
Nightshade: An offensive tool for artists against AI art generators原评论 ↗
引用causal · 2024-02

“Producing memorized images has precedent[0].”

指出生成被记住的图像已有先例,扩散模型会复述训练图像
Stable Diffusion 3原评论 ↗
引用zizee · 2025-07

“But Ai model have been shown to reproduce the training data”

引它说明 AI 模型已被证明能复现训练数据
Meta says it won't sign Europe AI agreement原评论 ↗
被提交为 story · 2 次