search

megvii-research / CADDM

Official implementation of ID-unaware Deepfake Detection Model
Apache License 2.0
135 stars 19 forks source link

about SBI #19

Closed fubenzun closed 10 months ago

fubenzun commented 11 months ago

I have a question，SBI AUC score is 93.18 in Celeb-DF dataset，Have you had a successful reproduction?I ran SBI official train code，only got 91.3。

Nku-cs-dsc commented 10 months ago

Hi, the SBI AUC score in our paper is quoted from the experiment part of the SBI paper. We conduct the comparative experiment based on adding the SBI augmentations in the data preprocessing method.

fubenzun commented 10 months ago

Hi, the SBI AUC score in our paper is quoted from the experiment part of the SBI paper. We conduct the comparative experiment based on adding the SBI augmentations in the data preprocessing method. HI. Have you re-run the experiments in SBI by official code？I ran the code only got 91.3 in CDF dataset,but the score in paper is 93.18?

Nku-cs-dsc commented 10 months ago

Sorry, we only reproduce SBI based on our code and never re-run the experiments with SBI official code.

cygbbhx commented 3 months ago

Hi, the SBI AUC score in our paper is quoted from the experiment part of the SBI paper. We conduct the comparative experiment based on adding the SBI augmentations in the data preprocessing method.

Hi, I am trying to reproduce CADDM + SBI experiment and have a short question about "adding the SBI augmentations".

  1. Did you use the self-blending technique 1) before you do the multi-scale swapping? Or 2) afterwards? or 3) randomly do SBI besides global/local swap?

  2. SBI uses only real images to generate own fake data. Did you also do so when running comparative experiments?