Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Just got doxxed to within 15 miles by a vision model, from only a single photo

twitter.com

Post reply on HN
Print view View original

31–40 of 67 posts

Re: Just got doxxed to within 15 miles by a vision model, from only a single photo

#31

Post by maxglute »

There's been a few examples where model guessed the city right because they see [insert street name] but the streetname is nowhere in the picture. But pictures had very definite tells like license plate style. I haven't gotten anything within 100km yet, mostly picking random cities in the correct country. Unless I include landmark, but feels like it's dropping random pins around the landmark because pinned location rarely gets the orientation / view correct. And it really should if the model references aerial maps, i.e. there's only 1 green space where the photo could feasibly be taken, but here's a random side walk.

Edit:

From the rainbolt interview, they claim trained on 200k images, 92% country accuracy with median error of 44 km... median error is... not remotely what I'm experiencing, but the AI seems to be doing great job in geoguesser match vs rainbolt. Also supposed trained to be aware of cardinal directions.

Re: Just got doxxed to within 15 miles by a vision model, from only a single photo

#32

Post by alberto_balsam »

echelon wrote: post #11
buildbot wrote: post #6

At the risk of sounding stereotypically skeptical, how do we know there's not a person behind the AI curtain, geo-guessing?

Rainbolt is one of the world's best GeoGuessr players: https://www.youtube.com/@rainbolttwo/featured AI repeatedly winning against him in a head-to-head: https://www.youtube.com/watch?v=ts5lPDV--cU

It's only mentioned at the end of the video but the AI is using features of the camera like locations of smudges. However it does appear to also heavily use genuine geographic fingerprints like telegraph poles. I wonder how it would fare against a human on a test set that is completely independent of the training set in terms of camera characteristics.

Re: Just got doxxed to within 15 miles by a vision model, from only a single photo

#33

Post by weinzierl »

shzhdbi09gv8ioi wrote: post #3

How well would it work if the scenery is not from a road mapped by Google Street View? Like in, the actual nature?

I tried it with a heavily distorted low-res screenshot from one of my Insta 360 videos of backcountry Portugal. You basically see a dirt road from above and a heavily distorted small strip of a hint of the ocean on the side - nothing more.

That it guessed this one correctly is a miracle to me.

Then I tried it with a photo of an abandoned structure in rural France and it guessed UK because some words in a visible Graffiti were in English.

So, it is hit and miss.

Re: Just got doxxed to within 15 miles by a vision model, from only a single photo

#34

Post by Toutouxc »

I tried three photos -- one from a hilltop in Prague (guessed Warsaw — 510 km away, but similar vibe), one from a hilltop in Bergamo (the guess was correct, it's a pretty touristy place) and one from a random street on Thassos, Greece (guessed mainland Greece). Nice stuff.

Re: Just got doxxed to within 15 miles by a vision model, from only a single photo

#35

Post by bdjsiqoocwk »

Of course you can dox from a picture of trees. You just need a database of where all the trees are.

People who believe "you can't dox from a picture of trees, because even tho conceptually trivial, you don't know how to search the database efficiently enough" are people who also do security thru obscurity.

Re: Just got doxxed to within 15 miles by a vision model, from only a single photo

#37

Post by dzhiurgis »

bdjsiqoocwk wrote: post #35

Of course you can dox from a picture of trees. You just need a database of where all the trees are. People who believe "you can't dox from a picture of trees, because even tho conceptually trivial, you don't know how to search the database efficiently enough" are people who also do security thru obscurity.

Sat pics could be enough to pinpoint a tree

Re: Just got doxxed to within 15 miles by a vision model, from only a single photo

#38

Post by magicalhippo »

Toutouxc wrote: post #34

I tried three photos -- one from a hilltop in Prague (guessed Warsaw — 510 km away, but similar vibe), one from a hilltop in Bergamo (the guess was correct, it's a pretty touristy place) and one from a random street on Thassos, Greece (guessed mainland Greece). Nice stuff.

Guess it depends heavily on training data. I uploaded an photo from nearby with a couple of very distinct towers, which I presume a good geoguesser would make short work of.

While it correctly guessed the country and general elements in the scene, it failed to pick up on those towers and the nuances of the terrain, which led it to suggest a very different place hundreds of km away.

Still, doesn't make me less wary of posting images from my neighborhood online... it'll only get better.

Re: Just got doxxed to within 15 miles by a vision model, from only a single photo

#39

Post by RGamma »

mastermedo wrote: post #21

Original link https://geospy.ai/ . Pretty scary stuff, I took a picture of one tree, got pinpointed to 4 km.

Gotta wonder why you'd even develop that. This thing seems to have essentially no value for for anyone but bad actors?
Post reply on HN