3 ms·

A bit hard to believe that VLU's NeuralHashes aren't also passed through CSAM detection, when NeuralHashes were first announced to the world in the context of C

by btown 4y ago

A bit hard to believe that VLU's NeuralHashes aren't also passed through CSAM detection, when NeuralHashes were first announced to the world in the context of CSAM detection: https://www.apple.com/child-safety/pdf/CSAM_Detection_Technical_Summary.pdf https://www.apple.com/child-safety/pdf/CSAM_Detection_Techni... That said, to reiterate the OP, there's no evidence that VLU executes when using Finder or QuickLook alone.

mftb 4y ago
Are you sure? Because in a prior article on the same source your parent linked it would appear that it is being used with QuickLook[0]. "While VLU is taking place, the image being looked up is opened in the floating window of a QuickLook preview." [0]https://eclecticlight.co/2022/03/23/how-visual-look-up-works-in-detail-1-paintings-on-intel-and-m1-macs/ https://eclecticlight.co/2022/03/23/how-visual-look-up-works...
- btown 4y ago
  Hmm, there is some ambiguity in the author's writing style. I interpreted what you quoted above as "if you are in Safari.app or Preview.app and activate VLU explicitly, it displays the image in a QuickLook window." The author does say in https://eclecticlight.co/2023/01/18/is-apple-checking-images-we-view-in-the-finder/ https://eclecticlight.co/2023/01/18/is-apple-checking-images... that: > Although the original description given was ‘Finder browsing’, for some that might include the display of images as QuickLook Previews, by selecting the image and pressing the Spacebar... [The process triggered from this] is consistent with the briefer task used in Live Text, and quite different from VLU. There is thus no evidence of the generation of neural hashes or any search query by PegasusKit typical of the later stages of VLU... > Local images that are viewed in QuickLook Preview undergo normal analysis for Live Text, and text recognition where possible, but that doesn’t generate identifiers that could be uploaded to Apple’s servers.
  - mftb 4y ago
    Understood. My own personal take-away, is that even with all the digging those folks have done it is not possible to know what Apple is actually doing. Which kind of reinforces your original point in the comment that I was responding to.
ribosometronome 4y ago
In the comments on the main article, the author says: >It’s perhaps worth pointing out here that VLU is nothing like Apple’s proposals for CSAM recognition. The neural hashes used to look up images in VLU are ‘low resolution’. They’re good for distinguishing specific images of paintings, breeds of dog and cat, and so on. But – as you can read in my detailed account of the proposed method – they fall far short of what would have been required to detect CSAM. Another big difference is that initial CSAM detection was intended to be performed locally, using a database of ‘positive’ neural hashes stored locally. Apple recognised that looking them up online wasn’t feasible, but that’s exactly what VLU does, because it works differently, with a much larger database and forming less robust and more inaccurate matches. So VLU ≠ CSAM detection, not by a long way. I'm still not super sure it couldn't be misused, though. A tool that is good enough to be able to reliably detect things like dog breeds (as the author also points out) seems like it should reliably enough also do things like detect drug paraphernalia, pictures of tankman, banned religious symbols, or whatever thing that people were more worried about being oppressed by a potentially authoritarian government. That's also why it seems, to me, like the more worrying vector for an authoritarian government to force abuse wasn't the higher resolution neural hashes matched against known images but the ML labeling of images for searchability.

by btown 4y ago

mftb 4y ago

Are you sure? Because in a prior article on the same source your parent linked it would appear that it is being used with QuickLook[0]. "While VLU is taking place, the image being looked up is opened in the floating window of a QuickLook preview." [0]https://eclecticlight.co/2022/03/23/how-visual-look-up-works-in-detail-1-paintings-on-intel-and-m1-macs/ https://eclecticlight.co/2022/03/23/how-visual-look-up-works...

btown 4y ago
Hmm, there is some ambiguity in the author's writing style. I interpreted what you quoted above as "if you are in Safari.app or Preview.app and activate VLU explicitly, it displays the image in a QuickLook window." The author does say in https://eclecticlight.co/2023/01/18/is-apple-checking-images-we-view-in-the-finder/ https://eclecticlight.co/2023/01/18/is-apple-checking-images... that: > Although the original description given was ‘Finder browsing’, for some that might include the display of images as QuickLook Previews, by selecting the image and pressing the Spacebar... [The process triggered from this] is consistent with the briefer task used in Live Text, and quite different from VLU. There is thus no evidence of the generation of neural hashes or any search query by PegasusKit typical of the later stages of VLU... > Local images that are viewed in QuickLook Preview undergo normal analysis for Live Text, and text recognition where possible, but that doesn’t generate identifiers that could be uploaded to Apple’s servers.
- mftb 4y ago
  Understood. My own personal take-away, is that even with all the digging those folks have done it is not possible to know what Apple is actually doing. Which kind of reinforces your original point in the comment that I was responding to.

ribosometronome 4y ago

In the comments on the main article, the author says: >It’s perhaps worth pointing out here that VLU is nothing like Apple’s proposals for CSAM recognition. The neural hashes used to look up images in VLU are ‘low resolution’. They’re good for distinguishing specific images of paintings, breeds of dog and cat, and so on. But – as you can read in my detailed account of the proposed method – they fall far short of what would have been required to detect CSAM. Another big difference is that initial CSAM detection was intended to be performed locally, using a database of ‘positive’ neural hashes stored locally. Apple recognised that looking them up online wasn’t feasible, but that’s exactly what VLU does, because it works differently, with a much larger database and forming less robust and more inaccurate matches. So VLU ≠ CSAM detection, not by a long way. I'm still not super sure it couldn't be misused, though. A tool that is good enough to be able to reliably detect things like dog breeds (as the author also points out) seems like it should reliably enough also do things like detect drug paraphernalia, pictures of tankman, banned religious symbols, or whatever thing that people were more worried about being oppressed by a potentially authoritarian government. That's also why it seems, to me, like the more worrying vector for an authoritarian government to force abuse wasn't the higher resolution neural hashes matched against known images but the ML labeling of images for searchability.