The results, especially the high numbers stated in the news article (68% recall, 90% accuracy) are overestimated as their verification method (i.e., whether the LLM detected really the right account) come from matching veryfied accounts with a test set of anonymous accounts of which they knew the real name. They knew the real name bcs the persons had a public link to their LinkedIn in their “anonymous” profile (which was removed for the sake of testing wheter the LLm can match the two acfounts. That being said: a user who uses a pseudonym but links his/her account publically to a, say, LinkedIn account doesn’t really care about anonymity and might hand out many more ‘breadcrumbs’ to follow than a truly anonymous account.

But I still think that also in the case of a fully anonymous account, people can be fingerprinted and matched with non-anonymous identities due to language, style etc. by a LLM.

And it will falsely identify people at even greater scale, because it is an imprecise and buggy tool.

For those who don’t know, we’ve been living in a dystopia since the 2000s.

Is this the first step towards using local LLMs for anonymity? 🫠 Always rephrasing each sentence somewhat. Truly dystopian stuff

Have they tried doing this for Satoshi Nakamoto yet?

The bright side - they can also be used to mask pseudonymous users. Guess how.

I am so grateful for already having been paranoid about sharing anything identifying about me starting 15+ years ago.

I never uploaded a picture of myself. Never used my real name anywhere. I used different nicks for different branches of the Internet. A plethora of different email addresses etc.

People thought I was being overly careful and I probably missed a lot of things due to not using Whatsapp, Facebook, Instagram, Twitter, Snapchat but I can’t say I regretted it at any point.

From a Facebook post I made on February 17th:

There are giant AI data firms that promise they can go through massive troves of data and pull out general and specific information from them. Information that is actionable and accurate. Give it 6 million data points and it’ll find all the links and organize them for you and unmask hidden details that aren’t visible to the naked eye.

Not one of those companies is stepping up to go through the publicly released Epstein files.

As a registered Republican woman from Texas with five children and two dogs, let me just say that I am astonished!

Yeah. I got a hunch of that a while ago, while trying some “old” scenarios of de-anonymization we used to do by hand. Just asking questions and posting pictures got surprisingly accurate results. A single picture with (to me) no significant landmark could lead to localizing a specific part of a city, and that was using a local LLM with a relatively small model, running on a 16GB VRAM 4060Ti.

It is now time to remember fondly the time where the younger people were warned by older people to not post all their stuff online, not over-share, be cautious about strangers, etc. I’m not sure when we lost that, but oh boy, it’s a festival.

Kind of obvious. If you’re a highschool teacher and you used to be a photographer. You also volunteer as a fireman. You live in France. You have 2 daughters. In 2022 you asked about repairs on your honda civic.
All off this can be amassed from different posts on facebook or reddit. There’ll be just a few people that fit this profile.

I think this will only work with people narrating their lives on social media.

“Got coffee from my favorite Granier at La Rambla! Ready of new day of work designing hats for dogs”

“Me and Bobby heading to Madrid to see my friend Concepcion. Do you like his new hat?”

“Just got nominated for ‘best business-casual hat’ at this year’s Barkies! So proud”

And so on…

Because how are you going to de-anonymize some random ramblings about Linux and beans? Everyone likes Linux and beans.

I theorized about this a long time ago. pretty sure I’m basically fucked

Of course, another option is for people to dramatically curb their use of social media, or at a minimum, regularly delete posts after a set time threshold.

Deletion won’t deal with someone seriously-interested in harvesting stuff, because they can log it as it becomes available. And curbing use isn’t ideal.

I mentioned before the possibility of poisoning data, like, sporadically adding some incorrect information about oneself into one’s comments. Ideally something that doesn’t impact the meaning of the comments, but would cause a computer to associate one with someone else.

There are some other issues. My guess is that it’s probably possible to fingerprint someone to a substantial degree by the phrasing that they use. One mole in the counterintelligence portion of the FBI, Robert Hanssen, was found because on two occasions he used the unusual phrase “the purple-pissing Japanese”.

FBI investigators later made progress during an operation where they paid disaffected Russian intelligence officers to deliver information on moles. They paid $7 million to KGB agent Aleksander Shcherbakov[48] who had access to a file on “B”. While it did not contain Hanssen’s name, among the information was an audiotape of a July 21, 1986, conversation between “B” and KGB agent Aleksander Fefelov.[49] FBI agent Michael Waguespack recognized the voice in the tape, but could not remember who it was from. Rifling through the rest of the files, they found notes of the mole using a quote from George S. Patton’s speech to the Third Army about “the purple-pissing Japanese”.[50] FBI analyst Bob King remembered Hanssen using that same quote. Waguespack listened to the tape again and recognized the voice as Hanssen’s. With the mole finally identified, locations, dates, and cases were matched with Hanssen’s activities during the period. Two fingerprints collected from a trash bag in the file were analyzed and proved to be Hanssen’s.[51][52][53]

That might be defeated by passing text through something like an LLM to rewrite it. So, for example, to take a snippet of my above comment:

Respond with the following text rephrased sentence by sentence, concisely written as a British computer scientist might write it:

Deletion won’t deal with someone seriously-interested in harvesting stuff, because they can log it as it becomes available. And curbing use isn’t ideal.

I mentioned before the possibility of poisoning data, like, sporadically adding some incorrect information about oneself into one’s comments. Ideally something that doesn’t impact the meaning of the comments, but would cause a computer to associate one with someone else.

I get:

The deletion of data alone will not prevent a determined party from gathering information, as they may simply record the information as it becomes available prior to its deletion. Moreover, restricting usage is not an ideal solution to the problem at hand.

I previously mentioned the possibility of introducing deliberate errors or misinformation into one’s own data, such as periodically inserting inaccurate details about oneself within comments. The goal would be to include information that does not significantly alter the meaning of the comment, but which would cause automated systems to incorrectly associate that individual with another person.

That might work. One would have to check the comment to make sure that it doesn’t mangle the thing to the point that it is incorrect, but it might defeat profiling based on phrasing peculiarities of a given person, especially if many users used a similar “profile” for comment re-writing.

A second problem is that one’s interests are probably something of a fingerprint. It might be possible to use separate accounts related to separate interests — for example, instead of having one account, having an account per community or similar. That does undermine the ability to use reputation generated elsewhere (“Oh, user X has been providing helpful information for five years over in community X, so they’re likely to also be doing so in community Y”), which kind of degrades online communities, but it’s better than just dropping pseudonymity and going 4chan-style fully anonymous and completely losing reputation.