search

JeremyCJM / CCST

Official repo for the WACV 2023 paper: Federated Domain Generalization for Image Recognition via Cross-Client Style Transfer.
Creative Commons Zero v1.0 Universal
23 stars 1 forks source link

Image Reconstruction Implementation #8

Closed judydnguyen closed 2 months ago

judydnguyen commented 2 months ago

Dear Authors,

Thanks for your great paper and public source code. I have a little concern about the experiments with image reconstruction experiments. I wonder why in the file https://github.com/JeremyCJM/CCST/blob/main/reconstruct_img/imagenet_reconstruct.py, we don't use any style statistic (style stats) of a specific domain. The loss function is the MSE loss between the reconstructed image and its corresponding style (not using the statistic information of the client -- a domain). I would really appreciate it if you could provide better guidance on this experiment. Thanks!

JeremyCJM commented 2 months ago

Hi Judy, it is because we need ground truth (GT) image to compute reconstruction loss during training, however, the average statistics do not have a corresponding GT. Therefore, the natural way is to train on image and image feature stats pairs, and then during inference time, you can test if you can reconstruct something meaningful from an average stat (domain style).

judydnguyen commented 2 months ago

Thank you for getting back to me. Yeah -- it makes sense. It will take more effort to design the inference attack using the clients' stats. I will mark it as close.

chenrxi commented 2 months ago

Thank you for getting back to me. Yeah -- it makes sense. It will take more effort to design the inference attack using the clients' stats. I will mark it as close.

Hi Judy, Are you working on the inference attack using the clients' stats? Are there any corresponding manuscripts on this topic?