RepMix: Representation Mixing for Robust Attribution of Synthesized Images

This repo contains official code and datasets for the ECCV 2022 paper "RepMix: Representation Mixing for Robust Attribution of Synthesized Images".

Dependencies

We experimented with the following main libraries (other versions may still work):

pytorch == 1.8.1
torchvision == 0.9.1
imagenet-C (see below)
opencv-python >= 4.2.0
Pillow == 8.3.1
pytorch-lightning == 1.4.6
...

The full list of dependencies can be found at requirements.txt.

To install imagenet-C:

git clone https://github.com/hendrycks/robustness.git && cd robustness/ImageNet-C/imagenet_c/ && pip install -e .

We also provide a Dockerfile so that you can build a docker image yourself. Alternatively you can download our pre-built docker image at:

docker pull tuvbui/ganprov:v1

The Attribution88 benchmark

The full dataset can be downloaded here (30GB). It consists of 12000x8x11=1056000 images of 11 semantics and 8 sources (real + 7 GANs). The train, validation and test splits are also included in the tar file.

We also release the processed test set here (5.4GB).

Train and evaluate

To train the RepMix model:

python train.py -d /path/to/attribution88/directory -tl /path/to/train/split/train.csv -vl /path/to/validation/split/val.csv -o /output/directory

To test a model:

python test.py -d /path/to/attribution88/test/directory -l /path/to/test/split/test.csv -w /path/to/model/checkpoint/last.ckpt

Reference

@InProceedings{bui2022repmix,
  title = {RepMix: Representation Mixing for Robust Attribution of Synthesized Images},
  author = {Bui, Tu and Yu, Ning and Collomosse, John},
  booktitle = {Proc. ECCV},
  year = {2022}
}