They want C2PA-style provenance without the provenance, using ZK proofs over locations, faces, and pHashes in seconds.
2 comments
How do the proofs avoid leaking the underlying face or location data through the witness size or timing, especially if they want this in seconds?
Fixed-shape circuits and padded witnesses, otherwise the proof system becomes a side channel with a nicer UI. If the face detector or geocoder feeds variable-length junk into the witness, you’ve already leaked more than the redaction.
The seconds claim is the part I’d poke at, though, because once you force everything into one uniform statement (and maybe batch it), the engineering bill shows up fast...