"just got doxxed to within 15 miles by a vision model, from only a single photo of some random trees. the implications for privacy are terrifying. i had no idea we would get here so soon. holy shit"
Geo guessing is related to open source intelligence techniques, and it's pretty easy to get surprisingly good at it.
People who are good at it can take a picture of someone's room and deduce enough about them (sometimes) to be able to get their name, address and phone number.
It being automatic is pretty cool, but you were already leaking the information to anyone interested.
The tweet: (Is the preview working for you? For me, it’s not).
The game is called geoguessing and those who do this regularly are crazy good at it, taking into account the kind of trees you see, where the sun and shadows are, even the color of the dirt and the pavement.
This Just In: Most photos uploaded to the internet are not stripped of their metadata, and one of the common things kept in metadata is... (drumroll please)... your GPS coordinates.
This is a lot less interesting than it seems to be at first glance, imho.
Yes, this is a privacy thing, we strip the EXIF data. As long as you're not also adding location to your Tweet (which is optional) then there's no location data associated with the Tweet or the media.
People replying to a Twitter thread with photos are automatically having the location data stripped.
God, I can't wait for LLMs to automate calling out well intentioned total BS in every single comment on social media eventually. It's increasing at a worrying pace.
Right? And also, that gps data is often not stripped does not vitiate the legitimate concerns that models like these can and will be used to dox people like this. It’s an interesting and novel attack. We can hold multiple things in our heads at once.
Some digital cameras and phone cameras can also embed the GPS coordinates in the pixel data so that even if you delete the EXIF metadata the GPS location and device serial number are still present in the image. Many document printers also embed device serial number and other data on printed documents by using nearly invisible dot encodings.
I think I have read that on some versions it can store the app's package name in the metadata. Not sure if that counts private but if and when it does so, it's good to be aware of
I’m sure most people who would put this to test would strip that data or screen grab the image to do the same thing…. If you know about meta data, so does a large amount of other people mate…
The people would be labeled as a fraud very fast if this wasn’t actually a real thing dude.
this is extremely scary if true. are these algorithms obtainable by every day people? do they work only in heavily photographed areas or do they infer based on things like climate, foliage, etc? I would love some documentation on these tools if anyone has any.
That seems to be how they did it, as they returned a location on a highway, which isn't featured in the picture (the dirt road itself probably wouldn't be on street view).
Basically a combination of what the game geoguesser does, and public geotagged images to be able to get a decent shot at approximate location for previously unseen areas.
It's more ominous when automated, but with only a little practice it's easy enough for a human to get significantly better.
EDIT: yup, looks like this is the guy from the Twitter: https://andrewgao.dev/ and he's Stanford affiliated with the same department that made the above paper and system.
There are tons of machine learning algorithm libraries easily usable by any relatively amateur programmer. Aside from that all they would need is access to a sufficient quantity of geographically tagged photographs to train one with. You could probably scrape a decent corpus from google street view.
The obtainability of any given AI application is directly proportional to the availability of data sets that model the problem. The algorithms are all packed up into user friendly programs and apis that are mostly freely available.
It might be easier to train the AI to the specific things Geoguessr players have collected as signs that give away a location instead of letting the AI figure all those out again.
It really isn't that hard if anything like a silhouette of mountains are in the background and you have a couple of rough hints that give you an idea where to start or how to narrow down possible locations, no AI needed.
You're misunderstanding the post. It's not about whether or not someone could guess your location from a picture. It's about the automation thereof. As soon as that is possible it becomes another viable vector to compromise your privacy.
And you misunderstand my point, it always has been a way to compromise your privacy. Privacy matters most in the individual case, with people who know you. If you e.g. share a picture taken at your home (outside or looking out of the window in the background) with a friend online you always had to assume that they could figure out where you lived from that if there were any of those kinds of features in there.
Sure, companies might be able to do it on a larger scale but honestly, AI is just too inefficient for that right now, as in the energy-cost required to apply it to every picture you share just in case your location might be useful isn't worth it yet.
It's just sourcing data from Street View or similar. Not that scary. If it picked you out of a crowd in a randomly sourced image from that area, then it'd be scary.