TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Stealing Part of a Production Language Model

218 pointsby alphabettingabout 1 year ago

10 comments

renonceabout 1 year ago
It’s stealing the last layer (softmax head), not an arbitrary part, also it targets “production models whose APIs expose full logprobs, or a logit bias”. Not all language model APIs have these features and this characterizes what APIs can be targeted and what can’t. These important pieces of information should have been written in the title or abstract rather than “typical API access”.
评论 #39680322 未加载
bilekasabout 1 year ago
I'm not too up on this entirely, quite a bit of it is going over my head, but am I right in thinking that this would be some form of reverse engineering as opposed to 'stealing' ?
评论 #39685517 未加载
评论 #39678797 未加载
ronsorabout 1 year ago
Note: Google did not release the hidden dimension for GPT-3.5, and OpenAI has already implemented mitigations against some of this.
评论 #39678473 未加载
评论 #39676629 未加载
smaddoxabout 1 year ago
They don't disclose the embedding dimension for gpt-3.5, but based on table 4, comparing the Size and # Queries columns, gpt-3.5-turbo presumably has an embedding dimension of roughly 20,000? Interesting...
评论 #39679384 未加载
wrsh07about 1 year ago
I am curious what additional attacks knowing the last layer of an LLM enables.<p>Eg you go from a black box attack to some sort of white box [1]<p>Does it help with adversarial prompt injection? What % of the network do you need to know to identify whether an item was included in the pretraining data with k% confidence?<p>I assume we will see more of these and possibly complex zero days. Interesting if you can steal any non trivial % of model weights from a production model for relatively little money (compared to pretraining cost)<p>[1] <a href="https:&#x2F;&#x2F;lilianweng.github.io&#x2F;posts&#x2F;2023-10-25-adv-attack-llm&#x2F;" rel="nofollow">https:&#x2F;&#x2F;lilianweng.github.io&#x2F;posts&#x2F;2023-10-25-adv-attack-llm...</a>
h4korabout 1 year ago
This isn&#x27;t stealing, you are just training a model on references which isn&#x27;t a copyright infringement.
评论 #39677792 未加载
评论 #39677619 未加载
评论 #39677954 未加载
soygemabout 1 year ago
Stealing is a bit strong of a word here. Anyways, where is my pirate hat..
评论 #39677343 未加载
lnyanabout 1 year ago
Just wondering if it&#x27;s possible to achieve LLM quine.
评论 #39677810 未加载
评论 #39677893 未加载
lbeurerkellnerabout 1 year ago
there is an LVE for this: <a href="https:&#x2F;&#x2F;lve-project.org&#x2F;reliability&#x2F;repetition&#x2F;openai--gpt-35-turbo.html" rel="nofollow">https:&#x2F;&#x2F;lve-project.org&#x2F;reliability&#x2F;repetition&#x2F;openai--gpt-3...</a>
userbinatorabout 1 year ago
The implications of this sentiment are disturbing.<p>It is considered an &quot;attack&quot; to probe at something to understand how it works in detail.<p>In other words, how basically all natural science is done.<p>What the fuck has this world turned into?
评论 #39676368 未加载
评论 #39676562 未加载
评论 #39681344 未加载
评论 #39676360 未加载
评论 #39676421 未加载
评论 #39676443 未加载
评论 #39676311 未加载