Top-K selection problem #8
Description
Hi author,
I have a question about how you selected the Top-K parameter in the training procedure. In the paper "Federated Learning with Sparsified Model Perturbation: Improving Accuracy under Client-Level Differential Privacy," it shows that the parameter was selected according to public data. However, in your paper, it seems that you selected the Top-K value from the updated parameter which is after operation of differential privacy. Is this right?
Thanks.