They train it on their private data, then summarize and encrypt the design’s new configuration. The design updates are sent back to the cloud, decrypted, averaged, and built-in into your centralized model. Iteration after iteration, the collaborative teaching continues right until the model is