← Back to context

Comment by google234123

4 years ago

How hard would it be to create a valid image that matches some 128bit hahs

If the details of the "hashing" scheme used is publicized, I imagine it will be near trivial. It's a long-standing problem in computer vision, to find a digital description of an image such that two similar images compare equal or at least similar.

State-of-the-art for this field is deep learning, and a /huge/ problem with the DL approach is that you can generate adversarial examples. So for example, a picture of a teacup that is identified by /most/ networks as a dog. It's particularly damning, because it seems like you don't have to do this for particular deep networks, they get tricked the same way, so to speak.

  • Since this algorithm presumably runs on-device, I am sure it won’t be long before someone has reverse engineered it…

    • Indeed, at which point we’ll know if Apple has implemented an obviously broken solution which opens us up to egregious government surveillance, or whether that is all just speculation without a factual basis.

If it’s a cryptographic hash - very hard.

  • This isn't cryptographic though. That would make the entire database absolutely trivial to bypass with tiny imperceptible random changes to the images.

    It's a perceptual hash.

  • It cannot be a cryptographically secure hash, simply because avoiding detection would then be trivial: change one channel in one pixel by one. Imperceptible change, different cryptographic hash.

  • But the probability is still not zero, and the number of iPhones in the world is large. A hash collision is possible, however unlikely.