Hacker News Reader

Comment by digitalsushi | Hacker News Reader

The big reason I don't want someone else scanning my phone (and a year later, my laptop) is that last I checked, software is not perfect, and I don't need the FBI getting called on me cause my browser cache smelled funny to some Apple PhD's machine-learning decision.

It's the same reason I don't want police having any opportunistic information about me - both of these have an agenda, and so innocent people get pulled into scenarios they shouldn't be when that information becomes the unlucky best-fit of the day.

That Apple has even suggested this is disturbing because I have to fully expect this is their vision for every device I am using.

And then I expect from there, it's a race to the bottom until we treat our laptops like they are all live microphones.

View on HN

The chance of you getting enough false positives to reach the flagging threshold is one in a trillion. And then the manual review would catch the error easily. The FBI won’t be called on you.

That's a claim by Apple. Due to the opaque process this is completely unverifiable.

As well this statement relies on the fact that no malicious actor out there is trying to thwart the system. The hashes used are deliberately chosen to easily produce collisions otherwise the system won't work. This almost certainly will be abused.

“This almost certainly will be abused.” is a claim by you and completely unverifiable. Apple has published white papers explaining how their hashing, matching, and cryptography work. How do you see someone thwarting that process specifically?

Apple's claims on how it will work are also completely unverifiable. What's stopping a government from providing Apple with hashes of any other sort of content they dislike?

And then Apple reviews the account and sees that what was flagged was not CSAM. And again, the hashes aren’t of arbitrary subject matter, they’re of specific images. Using that to police subject matter would be ludicrous.

How would Apple know what the content was that was flagged if all they are provided with is a list of hashes? I completely agree it's ludicrous, but there are plenty of countries that want that exact functionality.

The matched image’s decrypted security voucher includes a “visual derivative.” I’m sure in other words they can do some level of human comparison to verify that it is or is not a valid match.

If they have the hash/derivative they dont need to look on device or even decrypt, theyll know that data with this hash is on device, and presumably 100s of other matching hashes from the same device

Easy -- by having the authorities expand what is being searched for.

The concept of checking images against a hash list is not unique to Apple or even new.

https://en.m.wikipedia.org/wiki/PhotoDNA

What Apple announced is a way to do it on the client device instead of the server. That has some security implications, but they’re more specific than just “hashes might collide.”

They're not just checking against a hash list, they're using perceptual hashing, which is inexact and unlike cryptographic hashing or checksumming. Then, they use a fuzzy metric like hamming distance to determine if one image is a derivative of an illegal image.

The problem is that the space for false positives is huge when you use perceptual hashing, and it gets even larger when you start looking for derivative images, which they have to do otherwise criminals would just crop or shift color channels in order to bypass filters.

Manual review. So in other words they're going to go through your private photos unbeknownst to you to decide you're innocent. And you're never even going to know it happened. Wonderful.

What if it really was a very private photo?

> and then the manual review would catch the error easily.

How do we know this? It's not obvious to me that this system would work well for the edge cases that it seems intended for.

Reply on news.ycombinator.com