4 ms·

Can you elaborate how that worked? Apple only said contemporaneously that they would refuse to include hashes that aren’t CSAM, but never went into detail on ho

Can you elaborate how that worked? Apple only said contemporaneously that they would refuse to include hashes that aren’t CSAM, but never went into detail on how they could actually prevent that, or refuse an order to do so. But those hashes are produced and provided to Apple (edit: well, to any consumer of the hashes) by government backed actors. (Edited to replace “government agencies” with “government backed actors”.)

TheNewsIsHere 3y ago
I misread your reply, and I cannot edit the above. My misreading was that (paraphrased) “Apple could ensure the hashes they received were only for CSAM.” That isn’t what you said. You noted that (paraphrasing) “internally Apple would have some validation functionality to ensure they were not reporting non-CSAM content.” I was going to say I think your point is orthogonal, but I understand how it’s a response to what I wrote. I think we could rely on a reasonable set of people differentiating a perceptual hash of CSAM from non-CSAM content. That is no guarantee, but if their system worked as described then at some point they would be able to validate the actual stored images uploaded and know whether or not it was CSAM in an account. Just leaving here that I would counter to others that make your original point constrained to CSAM flagged content only, that: 1. Given the fact that collisions were demonstrated in public before the system was fully implemented, this failure mode had the potential to create an unknown volume of false positives that Apple would have to deal with (which would unnecessarily expose user data to Apple contractors or employees). 2. We probably don’t want to enable corporations to have this kind of power (see recent stories regarding similar false positives at Google). 3. There was proposed no detailed plans for any sort of procedures about what happened next. What happens when Apple makes a report that was “close to the line” and later found not to be CSAM? The perceptual hash concept was not without vectors for human failures. What recourse would there have been if Apple closed an account due to a false positive?

TheNewsIsHere 3y ago

I misread your reply, and I cannot edit the above. My misreading was that (paraphrased) “Apple could ensure the hashes they received were only for CSAM.” That isn’t what you said. You noted that (paraphrasing) “internally Apple would have some validation functionality to ensure they were not reporting non-CSAM content.” I was going to say I think your point is orthogonal, but I understand how it’s a response to what I wrote. I think we could rely on a reasonable set of people differentiating a perceptual hash of CSAM from non-CSAM content. That is no guarantee, but if their system worked as described then at some point they would be able to validate the actual stored images uploaded and know whether or not it was CSAM in an account. Just leaving here that I would counter to others that make your original point constrained to CSAM flagged content only, that: 1. Given the fact that collisions were demonstrated in public before the system was fully implemented, this failure mode had the potential to create an unknown volume of false positives that Apple would have to deal with (which would unnecessarily expose user data to Apple contractors or employees). 2. We probably don’t want to enable corporations to have this kind of power (see recent stories regarding similar false positives at Google). 3. There was proposed no detailed plans for any sort of procedures about what happened next. What happens when Apple makes a report that was “close to the line” and later found not to be CSAM? The perceptual hash concept was not without vectors for human failures. What recourse would there have been if Apple closed an account due to a false positive?