Parameter selection recommendations for the DPSGD algorithm

Any DPSGD algorithm in FashionMnist and cifar10 parameter selection suggestions? Such as sigma, C, learning rete.

Hey JeffffFu,

We generally advise to start with C=1 and the same learning rate as the vanilla, non-private baseline. Then, you can adapt sigma to the privacy/utility tradeoff you’re looking for.

Hope this helps,