News for hackers. Stuff that matters.
Re:Apple Delays Rollout of Child Safety Features by mlindner on Friday September 03, 2021 @11:15AM (#28405304)
I think it's actually impossible to use an AI to determine this type of thing. Even courts of humans sometimes rarely get it wrong sometimes with images that are open for debate (for example missing context) on whether it was abusive or not (parents bathing their kids, as one example). Then you have the whole can of worms that pedophiles will find sexual interest in images taken in non-abusive contexts but will use them none the less. How is an AI supposed to determine that. Even our laws can't handle that.

There's also images that can be abusive if made public but non-abusive if kept private. For example what about a kid taking pictures of themselves nude (many kids don't have parental moderated accounts) and putting them on icloud (iphone does this automatically with default settings).

It's a complete nuthouse and there's no way to do it.

(Also don't get me started on drawn artwork (or 3D CG) that people can create that can show all sorts of things and is fully or partially legal in many countries.)

Re:Apple Delays Rollout of Child Safety Features by scoopertrooper on Friday September 03, 2021 @12:04PM (#28405974)
They're not using an AI for that purpose. They're just sending a hash of the image and sending it along with iCloud uploads. The hashes are of images that have been identified by humans as containing child pornography.
Re:Apple Delays Rollout of Child Safety Features by eurasiantiger on Friday September 03, 2021 @01:57PM (#28407280)
No, the hashes are visual hashes. It is completely possible that a clean shaven adult taking pictures of themselves could get flagged.
Re:Apple Delays Rollout of Child Safety Features by mrtranscendence on Friday September 03, 2021 @03:18PM (#28408204)
It's also possible that a picture of an avocado will get flagged. A nude photo of someone who's shaved themselves is unlikely to look so much like an actual CSAM photo that it'll get flagged. It's certainly possible, but again, so's the avocado.
Re:Apple Delays Rollout of Child Safety Features by godelski on Friday September 03, 2021 @02:17PM (#28407502)
They're not hashes in the way md5sum or sha256 is a hash. They are _neural_ hashes. Which is not trying to fingerprint a specific image, but tag any image that is deemed sufficiently similar to the reference image. Hashes are unique, neural hashes aren't.

To add, there are natural neural hash collisions in ImageNet (a famous ML dataset). Images that look nothing alike to us humans.

Re:Apple Delays Rollout of Child Safety Features by Grustaf on Friday September 03, 2021 @02:31PM (#28407676)
Since the threshold is 30 images, and those images are matched to a curated database of abusive material, what you write here simply doesn't apply.