A research hospital called with a model that could spot a rare cancer two years early. The catch: it needed to train on ten thousand patient scans, and the hospital couldn't legally let a single one of them leave the building.

They tried federated learning. They tried synthetic data. Every path either leaked the patients or wrecked the model. Then they encrypted the scans and let the model learn from ciphertext it could never actually read.

the model never saw the faces

Fully homomorphic encryption meant the gradients were computed on encrypted pixels. The model got better at finding cancer, and the trainer never saw a single image it wasn't allowed to. That's the whole idea — the model learns, the data stays put.

privacy isn't a tax

For this hospital, privacy wasn't a compliance checkbox. It was the difference between having a model and not having one. Void didn't make it possible despite the privacy — the privacy was what made it possible.

— Daria NovakFounder, Void