LLMs can break online pseudonymity and identify users across platforms

General News
21

worked with tor. was checking it out. Noticing some strange issues and claims. What?

image
image996×571 36.9 KB

and

image
image1016×596 57.6 KB

22

Here are more details

I haven’t read it all but it seems to just be a bunch of “Why is Tor associated / working with Bad People X?” then legit concerns about node concentrations and some deanonymisation vectors.

But the leap to being a honeypot doesn’t make sense.

23

Seeing people call the Tor Network/Tor Browser compromised seems more like they’re trying to keep you from using it because it works.

1 Like
24

So just have an AI rephrase what you post. If you want to be pseudonymous.

We did it! We achieved dead internet theory. Bots talking to bots.

1 Like
25

This could quietly solve a lot of misinformation issues. If people spent just a few seconds drafting their response in a text document, reviewing it, and then realizing “no one wants to see that,” they could simply discard it.

26

Is using the same private LLM across a large number of users to convert text into a specific stylometry an effective defense against this attack?

The strategy of attempting to make many users’ fingerprints identical is employed by Tor Browser and Mullvad Browser, and since it has achieved some success, it appears at first glance to be applicable to text and seems effective.

27

Yeah I think it works, also you might want to be aware of your system prompt to reduce obvious artifacts in AI-generated texts.

1 Like
28

If a browser were to offer this as a privacy feature, I suspect Firefox would be the first to tackle it.
They have experience integrating local translation models into their browser and possess deep expertise in machine learning and anti-fingerprinting.

Personally, I think it would be difficult for Tor or Mullvad to implement it. They seem reluctant to adopt machine learning, even for local models.

1 Like
29

:laughing: I doubt it.

1 Like
30

It’s sad indeed. That being said,

I don’t think it’s foolproof, but a good solution. Your high-level info can still be deduced , like education, based on the meta-content.

31

I think your are not on the official instance: http://opbible7nans45sg33cbyeiwqmlp5fu7lklu6jd6f3mivrjeqadco5yd.onion/

32

it’s an okay website but it makes some strange claims as I posted above so..

33

Damn. Guess I have to delete every single message I have (including here) lol

Wonder if some platforms can protect it from LLMs