Innerworld

LLMs can unmask anonymous internet users for $1–4 each, matching 67% of pseudonymous Hacker News accounts to real LinkedIn profiles at 90% precision

https://lemmy.world/post/43902414

LLMs can unmask anonymous internet users for $1–4 each, matching 67% of pseudonymous Hacker News accounts to real LinkedIn profiles at 90% precision - Lemmy.World

Lemmy

Mar 6 at 12:47amWeb
Show threadtidderuufMar 6
Guess it’s a good thing I don’t use any social media with my real identity.
Show threadOnomatopoeiaMar 6

Right?

I have a linked in account which I haven’t touched in years, from a machine that no lonhers exists, on an internet connection I left behind.

Good luck connectinge to that.

Show threadinsufferableninjaMar 6
60% of the time it works every time
Show threadtestaccount372920Mar 6
67% of the time it works 90% of the time according to the article
Show threadRicky RigatoniMar 6
What does 67% at 90% precision mean
Show threadInnerworldMar 6
six seven
Show threadRicky RigatoniMar 6
Show threadlastaMar 6

Recall—that is, how many users were successfully deanonymized—was as high as 68 percent. Precision—meaning the rate of guesses that correctly identify the user—was up to 90 percent.

Show threadKairosMar 6

67% made a match. 90% of matches were right.

No idea how they got that number, though.

Show threadMadhuGururajanMar 6

Precision: ratio of true positives to total predicted positives.

Recall: ratio of true positives to actual positives

Show threadrefaloMar 6

No it can’t. This story keeps getting posted all over the internet.

Not only is it wrong, and not only do the researchers refuse to show their work (citing possible “misuse”), but it entirely depends on what kind of OPSEC failures the user happens to make.

Show threadsyddMar 6
So people without linkedin profiles are 100% safe?
Show threadEcco the dolphinMar 6

This headline sucks.

They made a model of accounts that willingly linked their hackernews profiles to their linked-ins and made a model base on that (n= approx 990)

They could “deanonymise” about 67% of those accounts from that n=990 candidate pool (alpha=.1) using their model (they already knew who they were, otherwise how could they verify a correct match?).

When they threw in a bunch of accounts that had nothing to do with those first accounts (89k total accounts) accuracy dropped to around 55%-45% depending on choice of technique.

  • first thing, those hn accts they trained on weren’t trying to be anonymous. They linked to their linked in profile. So, lie on the internet I guess

  • this is just a starting point anyway, cheap and fast. That’s what to worry about. $1-$4 per account you’re trying to doxx like this.

    • Just an interesting paper.