Open wl99123 opened 5 months ago
Hi, thank you for your interest. Note that this project is an unofficial implement for the paper "AdvDiffuser: Natural Adversarial Example Synthesis with Diffusion Models". And my reproduction results are not very satisfactory, because the synthesized samples look like they just have some noise points added. . The requirements are as followed: python_version == "3.8.16" pytorch~=2.2.0 torchvision~=0.15.2 matplotlib~=3.8.4 tqdm~=4.66.2 The chosen dataset is mnist, which is downloaded in main.py.
Hello, sorry to bother you again. I wanted to ask if you have done any experiments on ImageNet? Could you show me the effect of your synthetic images on MNIST if you have time?
Hi, sorry I just conducted experiment on MNIST. I've deleted the local folders, but I can recall that the synthesized samples look like they just have some noise points added. The official code page is https://github.com/lafeat/advdiffuser. They might have the figures of the experimental results.
What you wrote is very helpful for my work. Can you share your environment configuration and data design? Thank you