|
|
|
|
|
|
by sorryforthethro
2592 days ago
|
|
|
Federated Learning[1] you can train across multiple customer data by sending the in-training model to every device, then exfiltrating only the backprop adjustments you get back. In-theory, user data never leaves the device. In-practice, the backprop data has the information encoded within it to an extent, it's very difficult to secure. [1] https://medium.com/syncedreview/federated-learning-the-futur... |
|
|