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

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

twitter.com

65 pointsnotRobot62 comments

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

Took a picture and it recognized the architecture style, guessed the wrong town 60km away.

Same here: Picture from rural Austria in the North and it comes up with Hungaria. Location is in the center of Budapest, 300km distance to the original view point. But architectual style looks really hungarian.

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

Imagine: you’re abducted, but manage to sneak a photo and upload it somewhere completely random with a message for help. You don’t even know where you are.

Praise the lord, now I can leave my house without fear of being abducted.

It can do wonders for OSINT and disproving propaganda (for what's that worth).

Feels like there could be anti fake news applications. It's pretty common for photos to be recycled and purported to be of current events, when what they actually show is far away both in distance and time

location scouting for filmmakers

Interpol, Europol, FBI etc. find abused children by geolocating images. This could be very useful for that.

Just hire people from the geoguessr leaderboards, they are much better than this AI model. And every country employing like one or two such people and making them available to their police agencies is probably more than they could ever need.

You might think making an AI model will be cheaper, but if you commoditize it like this the additional police work from stalkers will outweigh any savings by multiple orders of magnitude

Um, no. Did a test and the results are neither scary nor wowing me now:

1) Easy mode: Cityscape of Naples, Italy; Bot 100% match, nice, I guess.

2) Mid-mode: A road in Glenoak Hills, CA; Bot says Malibu, CA... Never been there.

3) Hard-mode: Main lake in Ostróda, Poland; Bot says it's in Manhattan and that the vegetation matches Northern American :-D Ahahaha, didn't even get the continent right so...

Note: Tested on personal photos, with EXIF stripped, not widely/at all uploaded to the nets.

So... perhaps works great if you feed it back stock photos you've grabbed from the first result of a google search? Because my mileage is barely an inch.

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.

Does it matter? If this is something that can now be done cheaply then the effects of that are much the same whether it's an AI or not.

Mechanical turks aren’t news.

Maybe AI stands a for “Actually, Ivan!”

I know Rainbolt is handsome, but calling him a vision model maybe is bit of a stretch.

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.

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 some old phone photos from my grandparent's rural Canadian farm and it either got the province wrong or it guessed Montana or North Dakota.

there are services for photo sharing with location data (panoramio for example) that could be used for training.

Google Maps and Google Earth also lets you share.

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.

I gave it a selfie from the middle of some forest in Yellowstone and it nailed it. No signs or anything obvious, just a bunch of trees (also scrubbed EXIF).

I gave it a much harder photo from the Borrego Badlands, and it got SoCal, but was 200mi off.

When I show it random shots of mountainous jungle terrain from my patio (the view is long, maybe 20km in total…) if gets my location within 10-50km consistently.

I'm not convinced this is doing much - images with exif geo data seemed to just return the location literally described in the image. Images with no exif data returned every image location based on my IP location - a city a few 100 miles away from me, and even further from where the pictures were taken.

Some of them (pictures of a small town beach) were described as:

"The photo was taken from a boat in the XYZ Harbour, looking towards the city of <Large City Name>. The distinctive shape of the <literally tallest building in the country> Tower can be seen in the background. The ferry terminal is also visible on the left side of the photo."

needless to say, the tiny town in question is about 300km from the city mentioned, it does not have a single tower (I think it may be looking at a traffic pole?), and does not have a ferry terminal (there was a boat in the picture).

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.

Two photos (of rather distinctive places, same street), each one localized hundreds of kilometers away. At least the country matches.

I guess that this is really hit and miss depending on a bunch of parameters.

I wonder how Street View would perform on signatures such as Shazam uses

Sounds impressive but doesn't appear that great. Struck out three times on nature photos.

Chicagoland park-> guessed central park NYC

Rural Nebraska -> guessed Iowa

Shoshone Wyoming -> guessed Glacier Montana

Beyond the general look of places or particular constructed tells, I wonder if the model is taking any hints from analemma type patterns or latitude by the sun position (e.g. https://tvtropes.org/pmwiki/pmwiki.php/UsefulNotes/TheLongit... ). Cryptonomicon's use of that problem came to mind.

Posted a photo of my street, with Arthur's Seat in the background. (Meaning I'm in Edinburgh).

AI: Scottish residential houses, and a hill. This is clearly Glasgow.

Perhaps just trolling you?

This has been a thing for a while, pre LLMs:

https://www.reddit.com/r/wherewasthistaken

https://www.reddit.com/r/whereisthis

https://www.reddit.com/r/OSINT

I plugged in satelite view of pacific islands and it's giving random guesses. I used some of the source images listed on the bottom, one's with properly labelled island name, and it's still giving random guesses.

Random streetview shots seems to guess countries correctly so far, or at least region with similar construction. But that still means being (confidently) off by 1000s of km.

Satellite view? This is about street-level photos… I don't think you would mix these two things in the same model, it'd just dilute the capabilities on both.

It identified aerial shots of hawaiian islands, prince edward island, cuba and most of the commonly "known" islands with distinct shapes fine. At least much more consistently than random street level views I've thrown at it, where description seems sensible, in as much as a good geoguesser can descripe their rationales.

I don't question that some models can also handle satellite views, just that it makes little sense to "pollute" a model intended for geolocating photos with input data from a different category. Especially since satellite views are just vaguely similar enough to photos that I'd assume they can cause quite a bit of unnecessary ambiguity and model complexity to handle it.

> I don't think you would mix these two things in the same model, it'd just dilute the capabilities on both.

I'm not sure. Isn't the consensus that neural networks get better, if you make them bigger and give them harder problems?

> if you make them bigger and give them harder problems?

I mean… maybe… but that also makes them more expensive to run…

Speaking of Street View, the person who stalked a Japanese idol through her eye reflections [0] also used Street View.

[0] https://www.nbcnews.com/news/world/stalking-suspect-allegedl...

I tried a picture of open, expansive countryside in the UK midlands and it guessed the middle of Piccadilly Circus.

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

How do they know it’s not from EXIF data?

Most social media strips EXIF data to avoid that exact privacy problem.

That would probably be closer than 15 miles.

It's astounding that a vision model can do this.

When it comes to the fact that this is possible at all following various geolocation and OSINT accounts has had a quite sobering effect on me quite some times me ago.

I would like to point out josemonkey particularly, because he really explains the process very well. By and large I think it is first and foremost a fairly tedious task, but doesn't require any access to special services or special skills. Of course the vision model could change the tedious part.

Overall I think what the OSINT people do and what the vision model from the article simplifies is entertaining, but ultimately not super relevant, because reality is more like xkcd 538[1].

Most people have a device on them that in some way reveals their location more or less accurately. Location data in the grand scheme of things is a commodity for quite some time already.

While I think it is worth fighting that, I guess pragmatically we should all live under the assumption that our location is public information.

[1] https://xkcd.com/538/

Absolute rubbish in my case. Tried three locations with distinct landmarks, like mountains, lakes and buildings from Germany and Poland. In one case it got the country correct, in the others it guessed France.

This is indeed scary as hell. Tried it with a photo of the beach that could be any (sub)tropical beach in the world. Tt guesses Egypt and placed the marker about 100km away from the actual location.

Well that's horrifying

If you take unique photos, it's completely off. If you're taking boring photos that many other people have likely taken, it's very good.

Aren't there GeoGuesser types who can do this without a model? Some knowledge of vegetation goes a long way.

Yea but I guess AI makes it cheaper and more available. Just like photoshop pros could make fake photos before AI but now anyone can do it in 5 sec with a prompt.

Scalability and ease of use are the scary factors here. The barriers to getting involved with geoguessing were certainly a threshold for some people with bad intentions. These have now been removed or are considerably lower.

I wonder how many more capabilities the military info apparatus gained in the past few years.

Sounds terrifying

So, someone, probably a group, made the effort to build a system that can infer location from a picture. The negatives are obvious. What possible beneficial uses can it have?

Investigative journalism like the OSINT based researches conducted by bellingcat would be one example, as long as the AI results get doublechecked for accuracy.

CIA carveout

They could get geoguessers to participate, I guess. No need to build a publically accessible model.

Reply on news.ycombinator.com