4 ms·

Right, but the CSAM scanning routine is absurdly narrow and difficult to use for detecting anything but CSAM, whereas a general purpose hash matching algorithm

by zepto 5y ago

Right, but the CSAM scanning routine is absurdly narrow and difficult to use for detecting anything but CSAM, whereas a general purpose hash matching algorithm really is just a few lines of code. This whole ‘now that they have a scanner it’s easier’ reasoning doesn’t make sense. iOS already had numerous better matching algorithms if your goal is to do general spying.

Wowfunhappy 5y ago
Isn't the scanning routine searching for fingerprints? What happens if someone adds a fingerprint to the database, which matches something other than CSAM?
- zepto 5y ago
  No. It’s way more complex than that, and has been engineered to prevent exactly that scenario. I recommend you actually check out some of Apple’s material on this. The idea that it’s just a list or fingerprints or hashes that can easily be repurposed is simply wrong, but is the root of almost all of the complaints.
  - Wowfunhappy 5y ago
    I've looked through Apple's paper, what are you referring to?
    - zepto 5y ago
      So you know it’s not just a fingerprint. Did you notice the parts about the safety vouchers and the visual derivatives?
      - dannyw 5y ago
        Yes, I did. You do realise that CSAM is just a policy control, and there is literally nothing technical from Apple adding non-CSAM content to the same policy and systems? I feel like you may not have actually read the paper in depth. The paper clearly shows this is a generalised solution for detecting content that perceptually matches a database. The "what" is CSAM today, but nothing about the technicals require it to be CSAM. To answer your specific response of safety voucher and visual derivative: - The 'safety voucher' is an encrypted packet of information containing a part of the user's decryption keys, as well as a low resolution, grayscale version of the image (visual derivative). It is better described as "backdoor voucher". - The 'visual derivative' is just the image but smaller and grayscale. None of those two technologies have anything to do with CSAM. Apple's statement that they will only add CSAM to the database is a policy statement. Apple can literally change it overnight to detect copyrighted content, or Snowden documents, or whatever.
        zepto 5y ago
        You are ignoring how the parts of the technology are used to separate responsibilities. > Apple can literally change it overnight to detect copyrighted content, or Snowden documents, or whatever. No, Apple doesn’t run the database. NCMEC does.
        Wowfunhappy 5y ago
        It doesn’t matter where the fingerprints are coming from, they’re all just fingerprints. Today they come from NCMEC. Tomorrow they could be from the Chinese Communist Party.
        zepto 5y ago
        It’s hard to take seriously the complaint that this mechanism would give the Chinese Communist party more power. Also, in the US, the 4th Amendment protects against the abuse you are talking about: https://www.economist.com/united-states/2021/08/12/a-38-year-old-charity-will-be-integrated-into-apples-newest-operating-system https://www.economist.com/united-states/2021/08/12/a-38-year...
zionic 5y ago
This is false information, the scanning system is highly generalized to finding any type of photo via fuzzy matched perceptual hashes. It can target a pride flag just as easily as CP.
- zepto 5y ago
  No, what you have said is bullshit. The perceptual hashes are not generalized. They do not match ‘child porn’ or ‘pride flag’. They match specific photos from a database. You might argue that that database could contain a specific photo of a pride flag. You would be right, but there are safeguards against that. The system will not trigger with just one match. A set of multiple images must match. So no, the system won’t match a pride flag, and no the system won’t match a specific picture of a pride flag. The next argument would be that a whole load of pictures of pride flags could be uploaded and then someone who has enough of those pictures would trigger a match. This is true from a technical standpoint, but there are two more safeguards. The database is maintained by NCMEC, which is a non profit dedicated to tackling child abuse, and matches are visually confirmed by Apple. So, to match someone who just happens to have a pride flag in a picture is not possible. To match someone who has a set of specific pictures of pride flags is only possible if both Apple and NCMEC collude. So no - it can’t target a pride flag at all, much less as easily as CP.
  - simion314 5y ago
    >A set of multiple images must match. How many exactly? The threshold is secret, it could be 1 or 2 or maybe it is dynamic and depends on stuff like your identity.
    - zepto 5y ago
      It could be 1 or 2, but the claim is that it is set to only detect people who are actually building a collection of CSAM images because Apple doesn’t want this system to flag people who are unlikely to be actual criminals. I.e. they don’t just want no false positives, they don’t want positives that aren’t highly likely to be criminal. I’m guessing it’s a lot more than 2.

by zepto 5y ago

Wowfunhappy 5y ago

Isn't the scanning routine searching for fingerprints? What happens if someone adds a fingerprint to the database, which matches something other than CSAM?

zepto 5y ago
No. It’s way more complex than that, and has been engineered to prevent exactly that scenario. I recommend you actually check out some of Apple’s material on this. The idea that it’s just a list or fingerprints or hashes that can easily be repurposed is simply wrong, but is the root of almost all of the complaints.
- Wowfunhappy 5y ago
  I've looked through Apple's paper, what are you referring to?
  - zepto 5y ago
    So you know it’s not just a fingerprint. Did you notice the parts about the safety vouchers and the visual derivatives?
    - dannyw 5y ago
      Yes, I did. You do realise that CSAM is just a policy control, and there is literally nothing technical from Apple adding non-CSAM content to the same policy and systems? I feel like you may not have actually read the paper in depth. The paper clearly shows this is a generalised solution for detecting content that perceptually matches a database. The "what" is CSAM today, but nothing about the technicals require it to be CSAM. To answer your specific response of safety voucher and visual derivative: - The 'safety voucher' is an encrypted packet of information containing a part of the user's decryption keys, as well as a low resolution, grayscale version of the image (visual derivative). It is better described as "backdoor voucher". - The 'visual derivative' is just the image but smaller and grayscale. None of those two technologies have anything to do with CSAM. Apple's statement that they will only add CSAM to the database is a policy statement. Apple can literally change it overnight to detect copyrighted content, or Snowden documents, or whatever.
      - zepto 5y ago
        You are ignoring how the parts of the technology are used to separate responsibilities. > Apple can literally change it overnight to detect copyrighted content, or Snowden documents, or whatever. No, Apple doesn’t run the database. NCMEC does.
        Wowfunhappy 5y ago
        It doesn’t matter where the fingerprints are coming from, they’re all just fingerprints. Today they come from NCMEC. Tomorrow they could be from the Chinese Communist Party.
        zepto 5y ago
        It’s hard to take seriously the complaint that this mechanism would give the Chinese Communist party more power. Also, in the US, the 4th Amendment protects against the abuse you are talking about: https://www.economist.com/united-states/2021/08/12/a-38-year-old-charity-will-be-integrated-into-apples-newest-operating-system https://www.economist.com/united-states/2021/08/12/a-38-year...

zionic 5y ago

This is false information, the scanning system is highly generalized to finding any type of photo via fuzzy matched perceptual hashes. It can target a pride flag just as easily as CP.

zepto 5y ago
No, what you have said is bullshit. The perceptual hashes are not generalized. They do not match ‘child porn’ or ‘pride flag’. They match specific photos from a database. You might argue that that database could contain a specific photo of a pride flag. You would be right, but there are safeguards against that. The system will not trigger with just one match. A set of multiple images must match. So no, the system won’t match a pride flag, and no the system won’t match a specific picture of a pride flag. The next argument would be that a whole load of pictures of pride flags could be uploaded and then someone who has enough of those pictures would trigger a match. This is true from a technical standpoint, but there are two more safeguards. The database is maintained by NCMEC, which is a non profit dedicated to tackling child abuse, and matches are visually confirmed by Apple. So, to match someone who just happens to have a pride flag in a picture is not possible. To match someone who has a set of specific pictures of pride flags is only possible if both Apple and NCMEC collude. So no - it can’t target a pride flag at all, much less as easily as CP.
- simion314 5y ago
  >A set of multiple images must match. How many exactly? The threshold is secret, it could be 1 or 2 or maybe it is dynamic and depends on stuff like your identity.
  - zepto 5y ago
    It could be 1 or 2, but the claim is that it is set to only detect people who are actually building a collection of CSAM images because Apple doesn’t want this system to flag people who are unlikely to be actual criminals. I.e. they don’t just want no false positives, they don’t want positives that aren’t highly likely to be criminal. I’m guessing it’s a lot more than 2.