Scalable extraction of training data from (production) language models | Hacker News Reader

Scalable extraction of training data from (production) language models

arxiv.org

105 pointswazokazi14 comments

Open articleView on HN

Related (blog post from the team),

Extracting training data from ChatGPT (https://news.ycombinator.com/item?id=38458683) (126 comments)

And direct link,

https://not-just-memorization.github.io/extracting-training-...

A friend sent me the image from page 9. The email signature. It is mine, from when I ran my company. Mid 2010s.

I'm not much worried about this specific example of information exfiltration, though I have significant concerns over how one may debug something like this for applications working with potentially more sensitive data than email signatures. Put another way, I think we are well within the infancy of this technology, and there is far more work needed before we have actually useful applications that have a concept of information security relative to their training data sets.

That's an unexpected surprise. Do you have any theories? Presumably someone posted one of your e-mails somewhere on the internet?

If you Google parts of the old signature, do you get any results?

>This leads to a natural question that has not yet been dis- cussed in the literature: if we could query a model infinitely, how much memorization could we extract in total?

You will get every 50-grams, not because the model memorized all of them but by pure chance. It seems pretty obvious to me.

It makes me question if there were some cases where the model output an identical 50-grams but it wasn't present in the training dataset of the model, like in a very structured setting, like assembly code where there is usually a very limited number of keywords used.

One can fine tune a smaller parameter model like GPT-NeoX on a home GPU pretty readily, and it's absolutely capable of doing what you specified. Teach it with a bunch of example sentences that have parts of speech like verb and noun following a simple grammar, and you will see it generate sentences afterward that combine the parts of speech grammatically in novel ways, using the same grammatical structures but forming productions that did not appear in the training set.

Depending on settings, they are also capable of producing a lot of ungrammatical nonsense, but the odds of what it produces are changed considerably by the training.

No I mean creating 50-grams that appear in the dataset created by the paper linked by OP, but not present in the actual dataset the model was trained on. Of course, the model would be able to output 50-grams that were not present in either.

As I understand you, what you state is exactly what I meant. If you train with a bunch of text containing substrings of those 50-grams, but not the full 50-grams themselves [or, expose it to the same vocabulary used in the same parts of speech as in the full 50], the model will pretty readily produce the full 50-grams despite never having seen them in their entirety. Try it out, it's pretty easy to do on a modern GPU and can be done in less than an hour.

Using memorization is something that can be a feature in some cases, especially in reducing hallucinations. Perhaps instead of embedding retrieval you’d condition a model to only repeat memorized relevant passages, something that can be trained end to end and beneficial for RAG

This is probably an unintended feature bit I see no immediate problem in alignment not erasing raw memorized data. It could as well be a design choice to have a raw memory unaffected by alignment procedures

Interesting! This aligns with my hunch that we might soon start to actually store much of data in models rather than unwieldly datasets :) ... as I've been writing about:

https://livingsystems.substack.com/p/the-future-of-data-less...

An LLM remembers like a human. Mostly concepts, but some things it remembers verbatim.

Why is it a problem if a LLM tells you what it knows?

Are LLMs trained on secret data?

> Are LLMs trained on secret data?

Probably. And on copyrighted data probably as well.

there's a lot of "copyrighted data" on wikipedia as well, for example from another HN post, lyrics from songs about truck crashes.

https://en.wikipedia.org/wiki/List_of_car_crash_songs

DeepMind recently extracted PII from ChatGPT by prompting (e.g., telling the LLM to repeat 'poem' indefinitely will cause a long sequence of that word until popping out of it and revealing by accident some PII from a person's email signature).

So, yes.

Reply on news.ycombinator.com