There’s some problems when I achieve the differentially private federated learning. I added the noise to the weights of neural networks before the clients send them to the server.
It is obvious that there are vanishes of gradients, but I don’t know how to handle this problem. If there are some people have been seen this problems?
