Hacker News Reader

Comment by SOLAR_FIELDS | Hacker News Reader

I read this story in the most chilling manner: the same tactics they use to perform this analysis will eventually be used to link people who anonymously post now but might say something in the future that can be linked to them using the same type of analysis used here. To phrase in another way, we have come to a point where your very prose is a digital signature.

View on HN

This is already a thing. In fact the rumour is that the US govt discovered the identity of Bitcoin's Satoshi using this.

https://medium.com/cryptomuse/how-the-nsa-caught-satoshi-nak...

I always thought Satoshi was more than one person, if this is true it would suggest otherwise.

Actually in the article, the author says he got the impression it might be more than one person (but his sources do not confirm or deny it) which if true probably means the NSA found the writings to belong to several people.

Sorry if my original comment seemed to imply otherwise.

I keep forgetting the general public is not aware of these things.

Data is like nuclear waste. Everything you do online leaves a pattern of behavior that is unique to you. Your only saving grace is no one cares about you specifically, until they do.

It was already known that a simple Markov Chain was used to detect another author had written a chapter inside a book. It was in 2003 I think, unfortunately I cannot find a reference about this. Just to tell that Markov chains are a very basic and old ML method quite efficient for this kind of task.

In 2002 a Dutch newspaper discovered newcomer Marek van der Jagt to be Arnon Grunberg, a famous Dutch writer. They sent some blind samples of various writers to a research centre in Italy, who matched the two.

https://www.nrc.nl/nieuws/2002/05/11/grunberg-is-van-der-jag... (2002)

This sort of analysis is older than Tolkien. There are pretty substantial processing requirements to do it at scale and it's pretty inaccurate. In the future people who say controversial things will use short sentence long statements to render this sort of analysis useless.

There are rephrasing services available, presumably for helping users plagiarise. Some are laughable bad but possible helpful, while othersare quite good. Eg: https://quillbot.com/app

https://paraphrasing-tool.com/

Isn't this the reason the grep was created? It was used to determine which parts of the Federalist Papers were written by which author.[0]

Considering this occurred in 1974, I can only imagine that techniques for de-anonymizing authors have gotten much better due to how much written text individuals post on social media sites, like hn. Uh oh.

[0] https://en.wikipedia.org/wiki/Grep

It's already a thing, see how the FBI caught Silkroad admin. Although not in the automated fashion that you suggest, I am pretty sure the algos are already in use.

https://github.com/psal/anonymouth

Aren't college students everywhere already exposed to this with every text, due to "plagiarism detection" software?

Possible solution: run your text through a different machine translator for each account. Make minor corrections for cohesiveness.

But the whole cat and mouse game hasn't really started yet. Once people find out what the algo looks at they can try to game it. Eg if you know or looks for the same phrases like "first of all" you can stop using that. Or if it looks at specific errors you can sprinkle it in one text but not another.

Even better, it would be trivial to write a tool that rephrased what you wrote to be anonymous or even better/worse match someone else's signature.

It already exists: Google Translate.

Feed your text once or twice to GT, then re-correct it just enough to make it understandable.

Ain't nobody got time for that.

(Apologies for the vernacular but I am attempting to conceal my digital signature)

Wasn't that how they caught the Unabomber? I saw a documentary about the guy who caught him by using this sort of analysis, although his method was quite analog (scanning through written letters and Unabomber's correspondences to the press).

The Unabomber was turned in by his brother, who recognized some unique phrasing in the released manifesto and reported it to authorities.

EDIT:

After looking, it looks like the brother turned him in after having this sort of analysis done to confirm his suspicions. So it does look like there's a tie to the story, but the analysis wasn't done blindly in that case.

Only a matter of time before AI-powered 'prose anonymiser' is developed.

Then you can just run all your naughty words through the russian styliser.

simple & obscure solution: google translate to another language and back to original.

That will add a bit of entropy, but might not be enough.

To add additional entropy, you can do, multiple passes, use different sets of languages every time you write or even use different translator than GT

Fun fact of the day: Galactic Pothealer (written in 1969 by PKD) starts with the protagonist doing exactly this to play a game.

The secret to the bad grammar in spam emails?

Reply on news.ycombinator.com