Hacker News Reader

Comment by cholantesh | Hacker News Reader

I don't see how that makes it better.

Presumably, you would do this after Facebook already has a copy of the nude -- uploaded by the attacker/malicious ex.

If you want Facebook's computers to automatically take down any copies of the photo, yes, you're going to have to trust them with a copy of it.

Automated DMCA takedowns of copyrighted music and movies work in very much the same way.

> yes, you're going to have to trust them with a copy of it.

No. They could have you compute a fingerprint locally without uploading the image itself, preferably through some audited open source software, without auto-updates. They only don't do that because they want to guard their image hashing as corporate secrets.

No, Facebook's side needs to verify that the image is you, and that you're not just hashing a photo of the McDonald's logo.

they can verify that after they have found a match.

if there is no match you gave them no data of interest. if there is a match then they already have that image anyway and you didn't make your privacy situation worse, except maybe telling them that you claim that is you, now allowing them to associate more images was you, but that's probably an acceptable tradeoff if there is actual revenge porn of you out there.

It's not a simple hash. From Alex Stamos: “There are algorithms that can be used to create a fingerprint of a photo/video that is resilient to simple transforms like resizing.” That doesn't make it clear that they are making use of the powerful classifiers that Facebook has, but it's clearly not md5, either.

Regardless, I think your point stands.

They could also combine locally computed perceptual hashes of the unwanted images with face recognition of already uploaded regular images on the profile. That combination makes it even less necessary to send nudes to facebook.

My guess would be that they trained an autoencoder network for facial recognition and they use the encoder part for fingerprinting and check images by the decoder part.

https://en.wikipedia.org/wiki/Autoencoder

If they already have the image, then you don't need this system.

If they don't have it, but are doing verify-after, then the image won't actually be blocked right away. The proposed system would be of pretty marginal value. Some value, yes, but low enough that facebook decided it wasn't worth building that system.

You can't avoid having a tradeoff somewhere.

Presumably it’s both; hashing like this is fairly trivial to break if you can access the code.

If the recognition is done by autoencoder network as I suspect, then generative adversarial network could break it easily.

> compute a fingerprint locally

one downside: i think this could allow a bad actor to directly observe the fingerprinting and assist in developing a method to defeat it.

Or the attacker could just post the image elsewhere.

Or presumably an attacker could just slightly modify the image, and trick whatever CNN Facebook is trying to use into thinking it's not the same photo.

…a process which would be much easier given a local copy of the fingerprinting algorithm.

Trusted third parties.

That's what is missing from the internet, generally speaking. Services like public notaries. That's kindof what CAs provide. But there are a lot of similar services that could exist for things like this.

This is aimed at non-technical users. If we're lucky, maybe like 1% of users will be able to figure out how to do that. Not to mention all of the other issues with it mentioned below.

Well in that case it actually makes sense. They can fingerprint the photo and then automatically remove it.

Though really would be better to fingerprint all removed photos and automatically prevent any matching photos from being uploaded.

Do you have a better idea? I don’t.

Hash your copy of the nude image client side, and upload the hash only? Seems like a way better idea to me.

A hash would only match an identical file, right? It sounds like image recognition is used here to catch derivatives of the original file, which a hash wouldn't catch.

There are specific (non-cryptographic) hashes which are built to compare equal for equal data, and to be similar for similar data (for different metrics of similarity). It's a somewhat esoteric subject, but it's a thing.

The "hash" would have to be non-reversible. When we compare images based on content, afaik we transfer them into a vector space and measure distance. Do we really have tech to make and compare such a content representation in a non-reversible way? That sounds like a lot to ask for tech that seems still in active development.

All hashes are non-reversible. Being one way is one of the characteristics of a hash. Depending on how effective it is, there may be collisions possible. This is kind of happened with previously common hashes like md5 where people were able to break it by generating collisions.

But basically saying “as long as it is unreversable” is redundant unless you think there’s some reason to think the hash sucks.

If your hash is smaller than the (compressed) image itself, then you have information-theoretical loss; there is no possible way to reverse the operation. And these hashes are usually much smaller than the original.

I think it would be easy to create multiple hashes (reverse, flipped, etc...) and upload them all and look for matches. Provide the users with image recognition signature creating software (which is exactly what they are going to do) and let the users do it, vs uploading extremely personal and potentially embarrassing images to strangers on the internet.

How about recompressed, resized, photoshopped, format-converted... a filter against trivial transformations like transposition is worthless.

tough problem sure, but sending nude photos to a morally questionable company is not remotely close to the best solution to this problem.

then don't send them, but don't complain about facebook allowing your nudes to be posted. I don't think Facebook is going to open source their image recognition algorithm just because you want it to.

You do have a choice, use the service or not. If I ever found out a nude picture of me was leaked, I'd definitely trust FB more than whoever took it and leaked it in the first place. The whole point of the service is pretty much: your picture is already or about to be out there.

If someone has your nudes and intends to share them I think blocking them only on Facebook and not anywhere else is not productive.

You should just post the nudes yourself IMHO and be done with it.

^High School (and below) sure would be interesting /s

I imagine for most people in most situations regarding their* unwanted explicit media being published, the opposite is true i.e. Facebook(social media) is the worst place for it to be

It's called a perceptual hash, and there is a lot of research in this area.

Good thing most people know how to make a hash.

Reply on news.ycombinator.com