arakotom / dp_swd Goto Github PK

Dear Author,

Thank you for your interesting work!

I see that you have given the code for reproducing the Domain adaptation results, but there seems lack of experiments on the generative model, especially on CelebA dataset.

Therefore, can we expect the code about reproducing the results on the CelebA dataset?

Hi,
In the paper, you show how to apply DP-SWD for domian adaptation in Algorithm 2. If I understand correctly, the differential privacy of this alg is achieved by line 10-13, right?
I totally agree the distance computed by line 12 satisfy DP but I don't think the gradients computed by line 13 still satisfy DP because according to the chain rule, computing gradients of weights in hidden layers needs to use feature maps of hidden layers. And you do nothing to hidden feature maps, which means they are just non-linear maps of original private data, which can still exposure privacy information.
How do think about this?