Large-Scale Online Deanonymization with LLMs

The paper shows deanonymization from public posts. Imagine what's possible with private API traffic: the questions you ask, the code you paste, the errors you debug. Even if providers don't read it today, the data exists and the cost of analyzing it is going to zero.

Air-gapped local inference isn't paranoia. It's necessary.

Your interests can show up in all sorts of ways. Perhaps it's not saying "I like Madonna" on some social network, but the urge to interact with one specific song she recorded. One like can be the difference of giving away who you are or not.

With AI, there's a higher chance of active deanonymization tactics. This was possible for only select targets in the past. It's the creation of content or design of interactions that is meant to surface certain behavioral patterns (such as offering you that song "casually" in some timeline to gauge if you're going to interact with it).

Trying to mask or change your behavior is likely to result in a weird and very noticeable presence. Like trying to change how you walk will often lead to a caricaturized behavior, not something that someone would naturally do.

Acting naturally is probably the starting point of any attempt to prevent deanonymization, and the hardest to achieve. You have to be aware of your own behavior much more than people often do.

Step 1 was to scrape all of their posts into a database.

Step 2 was to have a human analyst review all of the posts for clues about who that person was

It was amazing that you could easily figure out:

- if they were at work or home from when they posted (9am to 5pm vs 6pm to 1am)

- what city they were in (based on sports teams, mentioning local landmarks etc0

- roughly what career they had

- their age based on cultural references

and mostly b/c they would drop a crumb of information here and there over months. They probably forgot about all of these individual events but when reading all of the posts in a few hours, the details became pretty evident. You get enough of these details and you can start to venn diagram people down to a few 100 likely candidates and then use LexisNexus style tools to narrow it down even further.

Given the above, it doesn't surprise me that LLMs can do the same but at high speed and across multiple sites etc.

My values or priorities may significantly change over decades, especially as a child, so why would I want to jeopardize the reputation of a potential future identity with something I may post today?