> Your phone will run your images through NeuralHash, and compare to the list of forbidden hash codes. If there’s a match, some data will be uploaded to Apple’s servers when you upload the image.
Your phone doesn’t know whether there is a match or not. There is no “if match then upload”. The safety voucher is attached to every iCloud upload, not just the matches.
This is why Apple can’t just flip a setting and scan all your images regardless of whether you add them to iCloud or not. If they wanted to run this against all your images, they’d have to either upload all your images to iCloud (easy to spot; difficult for them to do without consent) or change the way this mechanism works (not “just flipping a setting”; if they wanted to scan all your images, then why build it this way in the first place?).
1. https://cases.justia.com/federal/appellate-courts/ca10/14-32...
> Representatives of multiple law enforcement agencies have offices in the NCMEC building, including the FBI, Department of Homeland Security, U.S. Marshals, U.S. Postal Inspection Service ...
Technically, but not substantively, true since they report to the NCMEC, a government-funded government-established “private nonprofit” that is delegated a whole bunch of jobs for government, including gathering info on CSAM and distributing it to law enforcement for DoJ and processing child return/access requests under the Hague Convention on the Civil Aspects of International Child Abduction for the Department of State.
They do report you to a private organization who is primarily funded from the US Government and has zero interest in not passing along any information about potential child abusers.
If you believe Apple, there will be a human review element. What sort of people would go after this kind of job? I can't imagine. And this group will also have a huge incentive to err on the side of reporting any potential abuse content.
Edit: A legitimate, almost definitely "wants to the the right thing" private organization in the US. Have we heard about who gets notified in other jurisdictions?
And why would the NCMEC give law enforcement info about people they know are innocent?
For unencrypted content uploaded to their servers. That makes it a "Them Problem". Apple's approach is to scan before it is uploaded (and like the article says, "for now" - this can change with only a configuration change for all we know). And this could be a per-jurisdiction option and again, no one would know.
You have a point that this is not completely novel problem but do we have actual data on the number of real vs. false positives on the "everyone but Apple" group? Matthew Green probably would know but he doesn't[0].
[0] https://twitter.com/matthew_d_green/status/14283829528580833...
The left side photos are all apparently kitten photos but what if it was a photo of the Nirvana baby and, as a human reviewer, you had no access to the right side photo because it was just a hash in an anonymous database?
After all, Youtube copyright strikes have never been abused thanks to the fact that there are human reviewers.
Also, no app has ever been incorrectly removed from the Apple App Store, for the same reason that there are human reviewers.
This is thanks to the fact that human reviewers never make mistakes, unlike the rest of humans.
What could go wrong if we adopt the same system for something that could potentially result in a criminal investigation?
Are you able to explain why this specific case would be different?
Or maybe, more directly, what do you think Apple employees reviewing these hits would do if those types of images started showing up in the triage list.
The photographers posted them publicly on Facebook and Twitter. That’s how we have them in the first place.
The problem wasn’t knowing who took the photos, it was knowing who was in the photos. And that needs facial detection, which this system can’t do.
However, my primary point is that there is no possibility of a technical control over the scope of the program. It’s a policy decision and is subject to explicit expansion or collusion/parallel construction.
The article compares scanning to facial/object/scene recognition, which it can’t do.
The article talks about making it easier to find who took a photo, which is not the problem the Jan 6 sleuths were trying to solve.
The article worries about China, but ignores that matchable photos must appear in multiple jurisdictions’ databases.
The article also ignores the several points of third-party auditing and several layers of human reviews.