519
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 25 Oct 2023
519 points (93.6% liked)
News
35754 readers
716 users here now
Welcome to the News community!
Rules:
1. Be civil
Attack the argument, not the person. No racism/sexism/bigotry. Good faith argumentation only. This includes accusing another user of being a bot or paid actor. Trolling is uncivil and is grounds for removal and/or a community ban. Do not respond to rule-breaking content; report it and move on.
2. All posts should contain a source (url) that is as reliable and unbiased as possible and must only contain one link.
Obvious biased sources will be removed at the mods’ discretion. Supporting links can be added in comments or posted separately but not to the post body. Sources may be checked for reliability using Wikipedia, MBFC, AdFontes, GroundNews, etc.
3. No bots, spam or self-promotion.
Only approved bots, which follow the guidelines for bots set by the instance, are allowed.
4. Post titles should be the same as the article used as source. Clickbait titles may be removed.
Posts which titles don’t match the source may be removed. If the site changed their headline, we may ask you to update the post title. Clickbait titles use hyperbolic language and do not accurately describe the article content. When necessary, post titles may be edited, clearly marked with [brackets], but may never be used to editorialize or comment on the content.
5. Only recent news is allowed.
Posts must be news from the most recent 30 days.
6. All posts must be news articles.
No opinion pieces, Listicles, editorials, videos, blogs, press releases, or celebrity gossip will be allowed. All posts will be judged on a case-by-case basis. Mods may use discretion to pre-approve videos or press releases from highly credible sources that provide unique, newsworthy content not available or possible in another format.
7. No duplicate posts.
If an article has already been posted, it will be removed. Different articles reporting on the same subject are permitted. If the post that matches your post is very old, we refer you to rule 5.
8. Misinformation is prohibited.
Misinformation / propaganda is strictly prohibited. Any comment or post containing or linking to misinformation will be removed. If you feel that your post has been removed in error, credible sources must be provided.
9. No link shorteners or news aggregators.
All posts must link to original article sources. You may include archival links in the post description. News aggregators such as Yahoo, Google, Hacker News, etc. should be avoided in favor of the original source link. Newswire services such as AP, Reuters, or AFP, are frequently republished and may be shared from other credible sources.
10. Don't copy entire article in your post body
For copyright reasons, you are not allowed to copy an entire article into your post body. This is an instance wide rule, that is strictly enforced in this community.
founded 2 years ago
MODERATORS
I'm a professional software developer with ML experience, albeit not an expert in ML specifically. It would obviously affect the literal value of the embeddings, but there's no chance it would have a qualitative effect on a reasonably performant model.
It would though and their paper shows as much. The thing many forget is that it isn't trained visually like us. Little input changes like this have a big impact.
Now eventually if everyone uses the same glazing method the training won't care but at the moment this is bespoke enough that it can't be trained well on it. It will always be an arms race though.
No, it wouldn't, and the paper shows no such thing. Nightshade isn't "Gaussian blur + sharpen." It's based on the use of a different diffusion model to perturb an image (with bounded difference in perceptual similarity) to minimize the distance of the embedding from that of an unrelated concept. It is mathematically optimized and highly specific to the prompt. The clever thing is that you don't need access to the actual original text-to-image feature extractor because of the transferability between models, and the surprising thing is how few poisoned samples are required to break a model.
Blur+Sharpen isn't what Nightshade is doing, it's an example of a passive defense technique that may mess up fine-tuned "invisible" attacks because they rely on making minimal changes to jump category, and that can often come in the form of pretty precise pixel changes. You may have seen past papers about making pandas classify as gibbons. They rely on introducing a noise mask that just makes the image look a little worse quality, but in total is enough to flip the category. They don't really define their perturbation method in this paper, but there's some tension between being "invisible" and being resilient to "invisible" corrections like suggested above.
Oh, blur+sharpen to mitigate Nightshade makes sense, yeah.