Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

>But this doesn’t seem to be a slippery slope, either. Any photos you’re worried about can trivially circumvent the hashing system: adjust their brightness. Boom, done.

I've been briefed about it. Apple says explicitly that this is not the case. They mention that cropping, transforming, or even desaturating an image won't result in a hash that's so different from the matching one. The similarity will be high enough to detect the photo anyway. They don't seem eager to explain in detail how that actually works, though.



The reason they don't explain how it works is that the CSAM hash matching is an algorithm from NCMEC, which was apparently originally designed by Microsoft.

If you want to learn about the algorithm/hashes, you have to, among other things, sign an NDA about it, so it's likely that Apple, legally, cannot go into detail about how it actually works.

Microsoft has a page on it here if you're interested: https://www.microsoft.com/en-us/photodna

For obvious reasons, they don't go into a whole lot of detail.

I've learned a lot about how the process works from this blog post: https://hackerfactor.com/blog/index.php?/archives/929-One-Ba...

He gets a few facts about Apple's approach wrong, but it's an incredibly interesting read.


Apple designed a new system, which is different than PhotoDNA - they call it neural hash, and the specifics are not disclosed.


crosschecking hashes is only possible if they use the „same“ implemantation for the hashes. it would be safe to assume that the implementation/algorythm is the same regardless of being called photodna/neuralhash.


Sorry, I edited that part out one minute after commenting. I realized it was a weakness in the argument.

Thank you for pointing that out though. I had no idea.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: