search

royg27 / PAG-ROB

[ICML 2023 - Oral] Do Perceptually Aligned Gradients Imply Robustness? -- Official Code Repository
5 stars 1 forks source link

Using CM, OI or NN instead of SBG #1

Open rrmhearts opened 1 year ago

rrmhearts commented 1 year ago

Roy, we met at ICML. A quick question for you:

It seems that the SBG weights are required in the code

data = torch.load('data/c10_sbg_data.pt')
labels = torch.load('data/c10_sbg_label.pt')

Is there a straight-forward way to configure it to use class mean technique? I do not see any configuration besides the SBG technique.

Does the code only support SBG? If SBG is the only method supported by the available code, what exactly is the c10_sbg_data file? It's not the images themselves (which are downloaded) but is the file noisy versions of the data for computing diffusion-based gradients?

royg27 commented 1 year ago

Hi Ryan,

Thanks for your interest in our work. The code supports only the SBG method as it is the top-performing one. Regarding the sbg data file, it consists of both the original clean images and the per-class score-based gradients, extracted from a diffusion model. Thus, in the CIFAR-10 case, our data file consists of 11 views for each image (the image itself and the 10 per class gradients).

rrmhearts commented 1 year ago

Is there any way I could get access to the class mean code, despite it not being as good?

royg27 commented 1 year ago

Hi, please write me a private message at royg27592@gmail.com and I will try to help you reproduce it.