search

IBM / BadDiffusion

Official repo to reproduce the paper "How to Backdoor Diffusion Models?" published at CVPR 2023
Apache License 2.0
75 stars 5 forks source link

Details about ANP #15

Open FORRYKIM opened 2 months ago

FORRYKIM commented 2 months ago

When I looked at the ANP code, I found that the input perturbation_budget seems to be limiting the range of parameters for the batch normalization layer. For example, if perturbation_budget=4, the range of parameters before the limit is not greater than -4 to 4. How does the absolute size of perturbation described in the paper manifest?

FrankCCCCC commented 1 month ago

Thank you for your comment. As described in equation 13, we restrict the $\bar{\delta}{k}^{(l)}$ and $\bar{\zeta}{k}^{(l)}$ to control the injected perturbation in the model weight, which is the absolute size of perturbation.