This is specifically what Apple is trying to avoid - they are intentionally pushing an environment where the user must be involved to get the keys, by way of their account password and/or other enforcement mechanisms designed to ensure only the real user can access such keys.
This is discussed in detail in their Apple Platform Security guide: https://help.apple.com/pdf/security/en_US/apple-platform-sec...
All of the facial recognition, object identification, etc is all done on-device for the same reason. By contrast Google can and does do this in the cloud - and there are Google servers that intentionally have access to decrypt your photos.
iCloud backup was previously a vector that bypassed this, however they also today announced they are fixing that: https://news.ycombinator.com/item?id=33897793 "Apple introduces end-to-end encryption for backups"
But there are many serious problems with it. I'll isolate one: an on-device content surveillance mechanism is a slippery slope to a bad, bad place.
As the saying from the 90s goes, "child porn [CSAM] is the root password to the US Constitution." It is bad enough that people are willing to suspend their better judgement to do something about it.
But after you already have the mechanism accepted an in place, it is far easier to add to it. Pick your boogeyman: you're giving all of them a lovely tool to address their desires. Dissent suppression and Winnie-the-Poo detection for Xi, Erdogan gets to sniff out Golem memes, and choose-your-own horror for the coming dictator of the US.
It is far harder to tell a sovereign, "we could easily do that, but will not" than "we don't have a mechanism to do that." And pretending that it won't happen doesn't pass the laugh test - we have seen this show many, many times. But if you want to argue, start by explaining how Apple's jumping to implement the 10-minute-max sharing limit shows how they'd stand up to China about this.
It would help to note Apple also today announced "Advanced Data Protection" in iCloud which closes the hole where iCloud Backups, iCloud Photos and various other bits of data were technically decryptable by Apple. They've closed that (but it's opt-in, to balance the average user losing all their photos against other users desire to be secure even if it means losing their data). Details: https://support.apple.com/en-us/HT202303#advanced
However even without the "Advanced Data Protection" what I said about not having a workflow for any Apple server to "normally" request both the keys and photo data is also still good security.
This is what Apple was trying to avoid. Scanning on iCloud also requires that Apple can see your photos.
If the scanning is done on device, Apple could encrypt the photos in the cloud so that they can't decrypt them at all. Neither could the authorities.
> There's no reason to use customer's CPU/battery against them.
The amount of processing ALL phones do for phones is crazy, adding a simple checksum calculation in the workflow does fuck-all to your battery life. Item detection alone is orders of magnitude more CPU/battery heavy.
Why does that dystopia require on device scanning? Why couldn't they just do it with an OS update today? It's not a reasonably slippery slope, given the actual mechanics of how the CSAM system was designed (perceptual image hashes, not access to arbitrary files)
> There's no reason to use customer's CPU/battery against them.
That's the better argument, but still not super strong. On-device scanning means you know and can verify what hashes are being scanned for, who is providing them, and when they change. Cloud scanning is a complete black box. None of us would know if Google was doing ad-hoc scans of particular users' photos at the behest of random LEOs.
See my comment here:
https://news.ycombinator.com/item?id=33903825
> None of us would know if Google was doing ad-hoc scans of particular users' photos at the behest of random LEOs.
Not your device, not your software. You should assume anything you upload unencrypted is scanned. This distinction was clearly voiced by the majority during the debacle of Apple's on-device scanning proposal. They basically said, "Scanning in the cloud is [choose one: fine, skeezy], but we draw the line at doing on-device scans. I don't want that software on my device."
That would be precisely as difficult to verify as verifying whether your Apple device is currently scanning your content.
Once the feature exists your local data might become accessible to a government warrant, which would make the iPhone the opposite of a privacy oriented device.
If it's only for iCloud uploaded data they can simply do the scanning there. There's no reason to use customer's CPU/battery against them.