1. Privacy preserving. R4/R3: the features maps might leak privacy; R1: privacy property has not been described

We will discuss the privacy concerns in our revision. Analyzing the degree of privacy leakage under our framework will be our future works. Please note that the sizes of features and logits are fixed by the CNN architecture design. We will explain communication more obviously in our revision. R2: subsampling can save communication cost." Our method definitely can support the cross-device setting. It is better to tailor the strategy for different models and optimization methods. All clients are divided into many groups. The optimizer state of each group can be maintained by uploading it to the server of GKT. Once the group ID is changed, the server then synchronizes the optimizer state to the clients in the new group. We will demonstrate this in our revision. We can use the edge server in the hierarchical topology to improve load balance. Please provide IID and non-IID experiments for ablation study" The word "diverge" is somewhat misleading.