Machine Learning Privacy with Fully Homomorphic Encryption

The paper for the FFHE: https://eprint.iacr.org/2016/870.pdf

tldr here is that Zama believes privacy for the model owner (not the user) can be satisfied by having a trusted 3rd party host the (unencrypted) model and simply apply FHE to user data (which is faster). And that's the their solution.