I highly doubt this tbh. Not that I'm a homomorphic researcher, but in ML I try to keep an eye on it. Everyone I've talked to has always said it is 100x to 1000x slower. Though there is hope since some have told me that they are all CPU based. I have no idea if you could write cuda kernels for this or not. But still, there's a pretty big hit. If you do find a citation I'd love to see it too.
I think what bugs me about that comment is "differential privacy" is a different type of privacy than homomorphic, so would indicate that they are using... differential privacy...
Minor edit: I did find some research papers by Apple mentioning Homomorphic encryption, but nothing to suggest they are using it.
I think what bugs me about that comment is "differential privacy" is a different type of privacy than homomorphic, so would indicate that they are using... differential privacy...
Minor edit: I did find some research papers by Apple mentioning Homomorphic encryption, but nothing to suggest they are using it.