From a quick read of the PDF, my impression is this Safety Voucher is more of an additional metadata generated before a photo is uploaded to iCloud Photo (as oppose to scan photo library to see if it matches and notify Apple narrative). The Safety Voucher is designed in a way that they could decide on the server-side if a photo is CSAM without requiring iCloud to have an ability to decrypt all photos. Safety Voucher comes in three layers:
- The first layer is generating a NeuralHash for a photo and derive a cryptographic header from it based on blind hash provisioned by Apple from a known CSAM NeuralHash (e.g. Apple pre-compute this table for each device). First layer is decryptable if the generated NeuralHash is a known CSAM. (Is it going to be a random data if NeuralHash doesn't exists in the table?)
- The second layer uses Secret Sharing[1] scheme (where a content is decryptable only if >n keys are present) using NeuralHash of a visual derivative as a secret share. Multiple visual derivative is generated for each photo, and certain amount of shares is required to decrypt the layer.
- After the second layer is decrypted, the file is readable.
One thing that doesn't makes sense is iCloud Photos are not currently End-to-End Encrypted, so Apple already have access to all photos. This scheme would only makes sense if they plan to introduce End-to-End Encryption for iCloud Photos (& iCloud Backups?) in the future.
In my opinion, if they were to enable E2EE for photos, this scheme would be slightly better for privacy than Apple having an ability to scan all photos on the cloud (e.g. they can't retroactively decrypt a photo that already has Safety Voucher uploaded to iCloud if the matching NeuralHash doesn't exists at the time of blind hash table generation), but significantly less than having E2EE and not having this system in the first place.
(As someone from a country that is known to abuse this sort of power, I'm not happy about this system existing.)