AI-based Identification of Location from Photos

It looks like the day is rapidly approaching when AI models may be able to determine location somewhat accurately from photos, see this story:
https://www.npr.org/2023/12/19/1219984002/artificial-intelligence-can-find-your-location-in-photos-worrying-privacy-expert

This type of model could be a great boon for biologists/iNaturalists. For instance, it could probably easily flag images that might have an incorrect location and exclude them from datasets.

The privacy implications are also pretty clear though. Particularly , it could allow tracing of people’s movements or identification of observation locations from the photos if they have enough background info. This might be of particular concern for observations of organisms threatened by poaching/collection. I’m not really an AI alarmist per se, and at present, the tech doesn’t seem to allow for super precise geolocation (roughly looks like it might ID to the same general area as iNat obscuration in some cases), so this probably isn’t a current concern.

But for sensitive observations, it might be a good, sensible precaution to only include pics with closeups of the organism. In the future, and with major companies building this out (instead of students for a class), I can imagine the models might be much more accurate and background info could conceivably be used to defeat obscuration. Of course, that’s potentially the case now if there are very recognizable landmarks and humans that can ID them too, and the precaution that for very sensitive stuff, it may be best to just not post is still relevant.

8 Likes

Yeah, I am not sure about this. While it could help observations with a false location, it could also give new users the incorrect location that they may chose. Not only that, but a lot of people would be concerned about privacy. However I am not an expert.

Just my two cents worth.

4 Likes

I’d have to count myself as one of the “AI alarmists”, but not in the case of iNat images. I just scrolled rapidly through the first few pages of images under the Explore tab a few minutes ago and could find just two out of nearly 200 images that had anything resembling enough landscape information to potentially ID the location, and I doubt that would be possible even with the best AI. Coupled with a location pin, yes, of course a precise location could be determined, but the info in the images alone would not be sufficient.

Moreover, typically on iNaturalist, observations which show a small image of a critter or plant in a wide view are almost invariably pointed at common and/or planted species uploaded by relatively novice iNatters. So from a sensitive species protection viewpoint, I don’t see a high risk, especially for any observation which is obscured…but I could be proved wrong.

6 Likes

it is already possible to discover the underlying coordinates of just about any of the obscured observations in the system.

i don’t think any AI or algorithm that figures out locations based on backgrounds in images is going to perform better than the methods already available.

things that really need 100% protection simply shouldn’t be posted in the system.

2 Likes

it is already possible to discover the underlying coordinates of just about any of the obscured observations in the system.

Assuming you are talking about capital-O Obscured observations and not just Public observations where the posted location has been manually changed in some way from the geotagged one, that is incredibly alarming and should be made extremely clear by iNat itself, at the very least to new users, so as to not mislead them.

It’s explained here: https://www.inaturalist.org/pages/help#geoprivacy

3 Likes

What is not explained there is why iNat Australia (for example) doesn’t respect the geoprivacy I set on my own observations. This came up in another thread the other day and I probably should have made a comment there because I can’t find it now. Clearly the underlying coordinates are in the database so they can never be truly hidden from people with access to the database, but that iNat AU staff can see my locations (based on comments there) should be clearly explained, I think. It doesn’t really matter because generally I only hide coordinates because it’s my house or I’m at a property where the owner has asked me to not make the location visible. I’ve always assumed that staff can see my coordinates by looking at the underlying database, but it seems that iNat AU can see the locations more easily. Why?

Edit: I should add that I think that something like this makes me want to change my affiliation and there should be a warning. I was quite shocked to find out that my affiliation affected something like this

i can’t find it right now – maybe it’s changed – but the Terms of Use or one of those similar pages used to have a general warning similar to what i gave you in the last sentence of my earlier post.

even though it’s not super satisfying, i think that kind of general warning is the best warning available. i don’t think trying to explain the problem extremely clearly is actually the right way to go in this case. the Help document tiwane pointed to above tries to provide a lot of detail, but honestly, i think what’s provided there actually confuses the situation even more – not because what’s written there is inaccurate exactly, but because it gives the impression that people can understand the actual problem(s), which, frankly, i don’t think 99.9% of people will really understand, even if explained in excruciating detail.

i don’t think this is right. there are two ways to set geoprivacy. if you set it manually yourself, i don’t think iNat Australia will get your true coordinates. it’s only when the geoprivacy is set at the taxon level when true coordinates can be sent to iNat Australia (or other institutions with whom iNat shares such information).

1 Like

Based on the link Tony gave and the conversation the other day (which I can’t find) I think that because my affiliation is with iNat Australia they can see my obscured coordinates

Geoprivacy: If you set geoprivacy on your observations there are three ways the restricted location information can be shared with others:

  • If you ‘trust’ an individual
  • If you ‘trust’ a project
  • If you join via or affiliate with an iNaturalist Network Node (e.g. iNaturalist Australia)

Carrie Seltzer said something similar in the thread that I can’t find

1 Like

hmmm… maybe you’re right then. the Privacy Policy seems to support what you’re saying:

We share Personal Information associated with Users’ registration and account (non-public location data from Observations You post, IP address, email address, etc.) with representatives from iNaturalist Network Members (see https://www.inaturalist.org/pages/network for information about the iNaturalist Network) only for Users who have chosen to affiliate with that specific Network Member as their primary site. iNaturalist Network Members are responsible for the secure storage and responsible use of the data, and develop their own criteria on how data may or may not be reshared for non-commercial research, conservation, and species management purposes. You can prevent this sharing by affiliating with iNaturalist.org in your account settings.

Here’s how location access to a network site works:

Network partners get an export of all observations in their region twice a year. This export includes unobscured locations of observations if the observation is obscured only by taxon geoprivacy, regardless of who made the observation. So for example, I’m not affiliated with iNat Australia but I went there last year. For any observation I made that’s obscured due only to taxon geoprivacy, Atlas of Living Australia (ALA) will get access to their true coordinates in that export. But the true location of any observation I manually obscured will not be included in the export because I’m not affiliated with iNat Australia.

If you are affiliated with iNat Australia, then the export to ALA will include the true locations of all your observations in the region, regardless of the obscuration method. So if you don’t want ALA to get those locations, you should affiliate with iNaturalist, not iNaturalist Australia.

I hope that helps.

It’s not a warning, but there’s an explanation in the text under the affiliation chooser in Account Settings:

4 Likes

out of curiosity, can I ask why this is something you find so concerning? The only people who would be accessing these true coordinates of yours would be researchers/conservationists requesting them from the ALA

3 Likes

I find it concerning because if I’m visiting a person’s house and they ask me to not reveal the location of their house or property I obscure the location as per their request. If it’s not obscured because I use iNaturalist Australia, and the wording is a bit strange, I have two choices: a) I don’t upload those kind of observations at all; or b) I change my affiliation to inaturalist.org because it seems to have different rules (which is the bit I don’t understand)

1 Like

I should add that I don’t mind if a bunch of iNat staff turn up at my house based on my obscured locations. It might even be fun, we could have a BBQ and a couple of beers. What I don’t understand is why there are different rules of disclosure based on my affiliation. E.g. from above “So if you don’t want ALA to get those locations, you should affiliate with iNaturalist, not iNaturalist Australia” (edit: it’s a question of honour: I told the people I would obscure the location and if it’s not really obscured then that makes me a liar). Maybe it’s just the wording being used. Of course they have to share the underlying data with ALA, and I have no problem with that, but the way I read it things are more private if not using an affiliate site. Perhaps I’m just reading it wrong

While I appreciate the link, I don’t think it exactly addresses what @pisum was saying, unless I’m misunderstanding (and I’m taking into account their additional reply here as well) – the link here does not say that you can retrieve the exact location data for any obscured observation, just that if the same user posts a non-obscured observation nearby, or does not crop the photo well enough to make it less obvious, it could be possible to guess.

My understanding of the part I quoted from pisum in my earlier post was that setting the geoprivacy to Obscure does not actually obfuscate any geotagging present in the uploaded photo’s metadata, which, if true, would be a huge problem. Is this the case?

that’s not correct. other users cannot get to photo metadata on obscured observations. i don’t think it’s a good idea to poke too deeply into this. i think it’s enough to say that if you want 100% protection on any given observation, you shouldn’t post it.

2 Likes

I’m sorry for the misunderstanding. This is not the case.

  • all photos have their metadata stripped upon upload (aside from basic things like size, etc).

  • for any photo attached to an observation that has obscured or private geoprivacy, the metadata listed on its photo page does not publicly reveal location or date/time information. For example, this photo of mine is attached to an obscured observation. You should not be able to see any location or date/time data on that page. I can, however, because it’s my photo.

4 Likes

The spooky results are in spaces where there’s an abundant corpus of geotagged photos, relying on the clumpiness of human photo-takers. I wouldn’t expect that corpus of tagged data to pop into existence for lightly trafficked areas.
For many obfuscation purposes a time delay is also strong protection. I wouldn’t worry much about how data leaked today may allow pinpointing the location in five years time - I would think in many cases that data is no stronger than reasoning from the topology and looking for the right habitat.

Side Note: One of the “games” I enjoy in my leisure time (as an ecologist and biogeographer) is to carefully view the habitat and landscapes in car commercials on TV, particulalry those filmed in my home state of Texas, and try to assign a very precise location to them. Even in the absence of famous landmarks, it’s not too difficult to guess the biogeographic province of most of them. Sometimes landscape features allow for a county-level ID or closer.

5 Likes

I recall a story of a rare bird reported in the UK once that was concluded to be a hoax because in the photo it was standing on American-style barbed wire (I have no idea if the story itself was true…).