Why would anyone save CSAM to their photo library?
Specifically, many criminal actors don't touch CSAM because it's wrong. But some of these criminal actors will happily abuse legal systems, e.g. SWATTing.
This is not correct. Hash collisions won’t match the visual derivative.
That said I think your statement is a bit too strong, but generally true. A hash collision is not going to inherently be visually confusing. However you claim that it is impossible for an image to be both visually confusing and a hash collision, which seems unlikely. The real question is going to be how much more effort it takes to do both.
Also, the information needed to create a full match simply is not available.
Unless you're relying on it being computationally infeasible, but I'm not sure we know enough to consider that true at this point. Usually when we make statements on those grounds we do so with substantial proof. I don't think we know enough to do so here. I'm not even sure how feasible it is when you throw DL into the mix.
No.
The visual derivative is just a resized, very-low-resolution version of the uploaded image. "Matching the visual derivative" is completely meaningless. The visual derivative is not matched against anything, and there is no "original" visual derivative to match against.
If enough signatures match, Apple employees can decrypt the visual derivatives, and see if these extremely low resolution images look to the naked eye like they could come from CSAM. If so, they alert the authorities.. Given a way to obtain hash collisions, generating non-CSAM images that pass the visual derivative inspection is completely trivial.
The exact details of the algorithm are not public, but based on the technical summary that Apple provided, it almost certainly goes something like this.
Your device generates a secret number X. This secret is split into multiple fragments using a sharing scheme. Your device uses this secret number every time you upload a photo to iCloud, as follows:
1. Your device hashes the photo using a (many-to-one, hence irreversible) perceptual hash.
2. Your device also generates a fixed-size low resolution version of your image (the "visual derivative"). The visual derivative is encrypted using the secret X.
3. Your device encrypts some of your personally identifying information (device ids, Apple account, phone number, etc.) using X.
4. The hash, the encrypted visual derivative, and the encrypted personally identifying information are combined into what Apple calls the "safety voucher". A fragment of your key is attached to the safety voucher, and the voucher is sent to Apple over the internet. The safety vouchers are sent in a "blinded" way (with another encryption key derived using a Private Set Intersection scheme detailed in the technical summary), so that Apple cannot link them to specific files, devices or user accounts unless there's a match.
5. Apple receives the safety voucher. If the hash in the received safety voucher matches that of known CSAM content in the government-provided hash database (as determined by the private set intersection scheme), the voucher is saved and stored by Apple, and the fragment of your secret key X is revealed and saved. (You'd assume that they filter out / discard your voucher if there's no match; but the technical summary doesn't explicitly confirm this; this means that they may store and use it in the future to run further scans).
6. If your account uploads a large number of matching vouchers, then Apple will gather enough fragments to reassemble your entire secret key X. Now that they know your secret key, they can use it to decrypt the "visual derivatives" stored in all your saved vouchers.
7. An Apple employee will then inspect the "visual derivatives", and if your photos look like CSAM (more precisely, this employee can't rule out by visual inspection that your photos are CSAM-related), they will proceed to use your secret key X (which they now know) to decrypt the personally revealing information contained in your safety voucher, and report you to the authorities.
Keep in mind that the employee looking at the visual derivative does not, and cannot, know what the original image is supposed to look like. The only judgment they get to make is whether the low-resolution visual derivative of your photo looks like it can plausibly be CSAM-related or not. Plainly speaking, they will check if a small, say 48x48 pixel, thumbnail of your photo looks vaguely like naked people or not.
The relevant parts are.
> but based on the technical summary that Apple provided, it almost certainly goes something like this.
It doesn’t go like that. You are simply wrong.
Probably a mistake to say things like this, when the public documentation contradicts you.
> The visual derivative is not matched against anything, and there is no "original" visual derivative to match against.
Bullshit.
Here is the relevant paragraph from Apple’s documentation:
“as an additional safeguard, the visual derivatives themselves are matched to the known CSAM database by a second, independent perceptual hash. This independent hash is chosen to reject the unlikely possi- bility that the match threshold was exceeded due to non-CSAM images that were ad- versarially perturbed to cause false NeuralHash matches against the on-device en- crypted CSAM database. If the CSAM finding is confirmed by this independent hash, the visual derivatives are provided to Apple human reviewers for final confirmation.”
https://www.apple.com/child-safety/pdf/Security_Threat_Model...
Is that enough to cause an investigation? Maybe, maybe not, but I wouldn't want it to be a risk.
While I've no doubt that there's a lot of "before and after" images (which are still technically CSAM even if they're not strictly child porn) and possibly many innocuous images, they would not have been flagged as "A1".
I'm sure there's probably still a few images flagged as A1 which shouldn't be in the database at all, but that number is going to be small. How many of these incorrectly flagged images are going to make their way into your photo library? One? Two?
You need 30 in order for your account to be flagged.
And what’s the likelihood that a human reviewer will see these 30 odd images and press the “yep it’s CSAM” button?
More likely as soon as Apple’s human review sees these oddball images, they’re going to investigate, mark those hashes as invalid, then contact their upstream data supplier who will fix their data and now those implausible images are now useless.
A lot of people may not know how to avoid malware. But I don’t think very many of them would be so inept as to accidentally long press on child porn and tap “Add to Photos”.
Fixed it for you.
The image to be saved doesn't have to be disturbing at all to trigger a hash collision.
The linked repo has code to modify an image to generate a hash collision with another unrelated image.
That's the whole point.
But OK, say someone sends you a sunset that fools the hasher. Then what? Of course one match won’t do anything, so you’d need to download however many matching sunsets. Then what? The Apple reviewer would see they’re sunsets and you’d challenge the flag saying they’re sunsets. And if somehow NCMEC got involved, they’d see they’re just sunsets. And if law enforcement got involved, they’d see they’re just sunsets.
These proofs of concept might seem interesting from a ML pov, but all they do is just highlight why Apple put so many layers of checking into this.
We all want privacy but it seems odd to try to DoS this, with high risk for yourself and very little to gain.
Might be useful when the system turns into mass political surveillance tho.
A real attack would be to take legal porn images and make them collide with illegal images, so when a human goes to review the scaled down derivative images, those images very well look like they could be CSAM. Since there are many of them, they'd get sent to law enforcement. Then law enforcement would raid the victim's home and take all of their electronic devices in order to determine if they can be charged with a crime or not.
And that's assuming someone develops a hash collision which doesn't substantially mangle the photograph like the example offered on Github.
Specifically, only images categorised as "A1" are being included in the hash set on iOS. The category definitions are:
A = prepubescent minor
B = pubescent minor
1 = sex act
2 = "lascivious exhibition"
Do we know that for sure?
Apple has changed their mind enough times in the last week and a half that I'm convinced they're in full on defensive "wing it and say whatever will get people off our backs!" mode.
You can't read the threat modeling PDF and conclude that it was run through the normal Apple document review process. It reads nothing like a standard Apple document - it reads like a bunch of sleep deprived people were told to whip it up and publish it.
But by fog of war I was thinking more like the victim already has some sleazy (though marginally legal) stuff on their computer, or a search led to a find of pot in their house, or they lied to try and get out of the rap, or perhaps the FBI offered them a deal and they took it because they saw no way out, or perhaps they were simply an unlikable individual who the jury took a dislike to.
Basically that things are not always clear cut, and they come out of the wrong side of things, in a situation created by Apple's surveillance.
Surveillance is surveillance. It's a bit more obnoxious that a CPU which I paid money for is being used to compute the hashes instead of some CPU in a server farm somewhere (which I indirectly paid for) but the outcome is the same. The risk of being SWAT-ed is the same.
A sufficiently advanced catfishing attack could probably take advantage of this to get someone raided and have all their electronics confiscated.
Just send someone a zip of photos and let them extract it...
If it was anything like the image used to demonstrate this technique on Github, it's unlikely that anyone would describe that sunset as "beautiful". They'd be more likely to describe it as "bugger, this JPEG file is corrupted."
It was quite literally less than 24h from "Oh, hey, I can collide this grey blob with a dog!" to "Hey, this thing that looks like cat hashes to the same thing as this dog!"
You really think this is going to end at this proof of concept stage?
Regardless, this whole thing is moot because there are two classifiers, only one of which has been made public. Before any matches can make it to human review, photos in decrypted vouchers have to pass the CSAM match against a second classifier that Apple keeps to itself.