Concrete ML: transform machine learning models into a homomorphic equivalent
github.comI've been following these fellows for a bit, and haven't seen them mentioned here. This is a really fascinating product. What do data scientists and ML engineers think of this sort of tech breaking into the market?
So the idea is to train a model on clear data but provide inferences on encrypted data?
I think it could be interesting? Take recent events regarding Roe vs Wade and the fear driving women to uninstall period tracking apps. The best solution is to store data locally, push the model to the users and do inferences on device and never upload anything that could be tracked. Could this tool be used to make a safe version where the model runs on a server?
Edit: I must be misinterpreting something because the way to make that work would be by running X_test through encryption (client side) before passing the encrypted values to predict (server side) then finally decrypting y_pred_fhe client side as well.