after “gmaxwell” This is amazing Books in a cute collision published This contribution That he won’t publish any more collisions because he suspects Apple may be using/abusing his results and because Apple will change the NeuralHash algorithm (now in iOS before version 15).

My own thoughts on this: At (a very rough estimate) 100 million CSAM images and a NeuralHashlength of 12 bytes each iDevice owner submits More than 1 GB memory.

And I wouldn’t be surprised if, with relatively little effort, you can reverse 30+ NeuralHashes from this database; At least with the current (published) protocol.

If I understand correctly, NeuralHashes will be encrypted in the database with a public key. This public key is then provided so that NeuralHashes computed from the photos on your iDevice can be encrypted with it, after which this result is searched in the database.

Presumably, iOS will also look for slight differences in NeuralHash. After all, NeuralHash (the current version) is an independent 96-bit vector (array). By flipping those 96 bits one by one (and encrypting the result with the public key and looking it up in the database) you can check with hammer distance From 1. It is clear that the 2 Hamming distance requires more calculations.

Anyway, 96 bits are weak by modern cryptographic standards (i.e. brute force might be possible). These are not pseudorandom numbers, but they are meant to be biased.

I suspect that if you typically start a brute force attack (porn) and then increase your Hamming distance, you will be able to find your unencrypted NeuralHashes much sooner than you might expect.

We’ll have to wait and see what Apple brings in iOS 15. On the other hand, they can’t put as many GB databases on 64GB smartphones, and in the meantime, the number of known CSAM images is only growing (not to mention you want to know about movies) .