ducadveritatem

2 lata temudiscuss

Based on their presentations yesterday and accompanying written materials they have a pretty capable ~3B parameter Foundational LLM running fully locally on the devices which is the first line. https://machinelearning.apple.com/research/introducing-apple...

ducadveritatem

2 lata temudiscuss

Good news then that you’re explicitly asked for permission each time any query would by shared with ChatGPT.

ducadveritatem

2 lata temudiscuss

You can read what they have shared about the security aspects of the cloud portion of their AI offering here: https://security.apple.com/blog/private-cloud-compute/

ducadveritatem

2 lata temudiscuss

Can’t find a link right now (search terms are pretty crowded atm!) but I saw they just recently shared an llm they’ve been working on that is designed to answer questions about how a give screen of an app functions. (Identifying buttons, core functionality, etc.)

ducadveritatem

2 lata temudiscuss

I actually liked that they didn’t show any of the AI writing capabilities being used in iMessage but rather in email client for more professional contexts. I’m really curious to see if they make it available in iMessage…

ducadveritatem

2 lata temudiscuss

You can only go into so much detail in that format for that audience. I was happy that they simultaneously posted a much more technical deep dive: https://security.apple.com/blog/private-cloud-compute/

ducadveritatem

2 lata temudiscuss

The details of their approach: https://security.apple.com/blog/private-cloud-compute/

ducadveritatem

2 lata temudiscuss

Word on the street (someone who was talking to Apple employees at WWDC) is that the Vision Pro doesn’t have enough headroom on the processor for it. It’s driving that sucker really hard just doing its “regular” thing.

ducadveritatem

2 lata temudiscuss

If you want the details: https://security.apple.com/blog/private-cloud-compute/

ducadveritatem

2 lata temudiscuss

Before you write off their claims I encourage you to read more about the detailed specifics (if you have the technical footing and inclination to do so). While the approaches should certainly be probed and audited, it’s clearly more than performative. https://security.apple.com/blog/private-cloud-compute/

Also a good thread from Matthew Green, a privacy/cryptography wonk, on the subject: https://x.com/matthew_d_green/status/1800291897245835616?s=4...

ducadveritatem

2 lata temudiscuss

What do you mean by the “record”? It seems like you think this means Apple somehow has access and stores all that information in their cloud and we just have to hope/trust that they don’t decide they want to poke around in it?

You should look more into their security architecture if you’re curious about stuff like this. The way Secure Enclave, E2EE (including the Advanced Data Protection feature for all iCloud data), etc. The reality is that they use a huge range of privacy enhancing approaches to minimize what data has to leave your device and how it can be used. For example the biometrics you mention are never outside the Secure Enclave in the chip on your phone and nobody except you can access them unless they have your passcode. Things like running facial recognition on your photos library is handled locally on your device with no information going up to the cloud. FindMy is also architected in a fully E2E encrypted way.

You can browse their hundreds of pages of security and privacy documentation via the table of contents here to look up any specific service or functionality you want to know more about: https://support.apple.com/guide/security/welcome/web

ducadveritatem

2 lata temudiscuss

You can read the details of their approach for the privacy/security aspects of the cloud compute portion here. https://security.apple.com/blog/private-cloud-compute/

ducadveritatem

2 lata temudiscuss

Are you claiming Apple uses personal user data (e.g someone’s photos or texts) as training data for their server-side models? That’s a massive claim and there are some journalists you should definitely shoot a message to on signal if you have proof of that and aren’t just blowing smoke.

ducadveritatem

5 lat temudiscuss

The explanation is complicated but really fascinating. I think I understand it, but not well enough to explain it. Read the section entitled "What cryptographic tools are used in the implementation of the system?" in this write up about Apple's methodology.

https://www.apple.com/child-safety/pdf/Technical_Assessment_...

Also look at their full whitepaper here: https://www.apple.com/child-safety/pdf/CSAM_Detection_Techni...

My take is that performing the initial hash matching and encrypting the results in two separate layers on device prevents Apple from having meaningful knowledge of low (under the set threshold for being flagged) quantities of matches on a user account. This protects the use of the threshold as a way to further reduce false positives. For example they couldn't comply with a subpoena that said "Hey, we know you set a threshold of only flagging + reporting accounts with 50 image matches, but we want to see a list of all accounts with 10 or more matches because we think that's good enough."

This method lets them set and enforce a threshold to maintain their target false positive rate which they say is ~1 in 1 trillion accounts incorrectly flagged.

Disclaimer: I'm not a cryptographer and could be misunderstanding this.

ducadveritatem

5 lat temudiscuss

Sorry, but they're not "spying on your device". If you choose to upload your photos to their cloud service (iCloud Photo Library), they'll use perceptual hashing and private set intersection tech to check against a database of known exploitative child imagery to make sure you aren't uploading that to their service. That's it.

ducadveritatem

5 lat temudiscuss

That's exactly what they're doing. And they're requiring multiple matches before an account is flagged, further reducing false positives. They claim they've set the threshold to achieve a ~1 in 1 trillion probability of an account being improperly flagged.

ducadveritatem

5 lat temudiscuss

The "algorithm" isn't some sort of neural network trying to intelligently identify things that "look like" CP. It's a perceptual hash matching against a database of known CP. It has to find multiple matches before it flags the account for review to reduce false positives. Only after review confirming a match to known exploitative images is the info referred to NCMEC for action.

Full whitepaper: https://www.apple.com/child-safety/pdf/CSAM_Detection_Techni...

ducadveritatem

5 lat temudiscuss

That isn't accurate. They're not blindly handing lists of users over to the government.

If an account uploads multiple images that match to known exploitative images and exceeds a threshold, then the account is flagged for review by Apple. (Note the threshold is selected to provide a ~1 in 1 trillion probability of incorrectly flagging an account.) Once they review and confirm a match, it's then forwarded to the National Center for Missing & Exploited Children for further action (and presumably referral to Law Enforcement.)

More details in their whitepaper: https://www.apple.com/child-safety/pdf/CSAM_Detection_Techni...

ducadveritatem

7 lat temudiscuss

Or maybe "don't work at Apple if you plan to spend dozens of hours of company time on the phone with your startup cofounders and use proprietary HR information to poach employees while still employed yourself."

ducadveritatem

7 lat temudiscuss

This just seems like them attempting to be sensationalist. Anyone who has a company device managed with MDM that doesn't think their employer can monitor everything that happens on that device should go back and reread the terms they accepted when they got the device.

comments