This repository is the implementation of the paper:

Dragoș Țânțaru, Elisabeta Oneață, Dan Oneață. Weakly-supervised deepfake localization in diffusion-generated images. WACV, 2024.

The code depends on Pytorch, which has to be installed separately. For example:

conda create -n dolos python=3.9
conda activate dolos
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.6 -c pytorch -c nvidia

Then run:

pip install -e .

This repository also integrates code from the following repositories:

• Patch Forensics
• Facial Forgery Detection

Our datasets are available here. They can be downloaded remotely as follows:

wget "https://sharing.speed.pub.ro/owncloud/index.php/s/8kPtPO8efjRMUAG/download" -O dolos-data.zip

After unzipping the dataset, with

unzip dolos-data.zip

set a symbolic link in the root of this repository; for example:

cd /path/to/dolos && ln -s /path/to/wacv_data data

Note. The masks are given as black-and-with images, where black (0) means fake and (white) 255 means real. However, the models below adopt the opposite convention: 0 means real and 1 means fake.

In the paper we consider three methods for weakly-supervised localization:

• Grad-CAM. Implemented in dolos/methods/grad_cam.
• Patches. Implemented in dolos/methods/patch_forensics.
• Attention. Implemented in dolos/methods/xception_attention. For training, we initialize from the pretrained model provided by the authors. The weights can be downloaded as follows:

mkdir assets
wget https://github.com/JStehouwer/FFD_CVPR2020/releases/download/v2.0/xcp_reg.tar -O assets/xcp_reg.tar

These three methods are evaluated on three experimental setups:

• Setup A: a weakly-supervised setup in which we have access to fully-generated images as fakes and, consequently, only image-level labels.
• Setup B: a weakly-supervised setup in which we have access to partially-manipulated images, but only with image-level labels (no localization information).
• Setup C: a fully-supervised setting in which we have access to ground-truth localization masks of partially-manipulated images.

The weakly-supervised setups (A and B) are implemented by the train_weak_supervision.py script, while the fully-supervised setup (C) is implemented by the train_full_supervision.py script. For example:

ls dolos/methods/xception_attention/train_weak_supervision.py
ls dolos/methods/xception_attention/train_full_supervision.py

In this set of experiments we evaluate the localization capabilities of the three methods in the three setups. The table below contains links to the scripts that can replicate the experiments in Table 2 from our paper:

For this setup, we train the Patches model on one source dataset (FFHQ) and evaluate it on another target dataset (CelebA-HQ); note that the generation method is kept fixed (P2 diffusion). Below are the scripts needed to obtain results for the three setups, corresponding to Table 3 in the paper:

• Setup A: run-patches-setup-a-ffhq.sh
• Setup B: run-patches-setup-b-ffhq.sh
• Setup C: run-patches-setup-c-ffhq.sh

Here we evaluate the Patches detection model across multiple combinations of train-test datasets. For these experiments, we vary the generation method, but keep the source dataset fixed (CelebA-HQ)

To train on a given dataset:

python dolos/methods/patch_forensics/train_full_supervision.py -c repaint-p2-9k

The datasets are specified by a configuration, and can be set to one of the following: repaint-p2-9k, repaint-ldm, lama, pluralistic, three-but-repaint-p2-9k, three-but-repaint-ldm, three-but-lama, three-but-pluralistic.

To obtain predictions for all combinations:

for src in repaint-p2-9k repaint-ldm lama pluralistic three-but-repaint-p2-9k three-but-repaint-ldm three-but-lama three-but-pluralistic; do
    for tgt in repaint-p2-9k repaint-ldm lama pluralistic; do
        for split in test; do
            python dolos/methods/patch_forensics/predict.py -s full -t ${src} -p ${tgt}-${split}
        done
    done
done

Finally, to obtain the evaluation metrics and generate Figure 6 in the paper:

streamlit run dolos/methods/patch_forensics/show_cross_model_performance.py

We provide the weights of the Patches method trained in the setups described in sections 5.1 and 5.2. The models are available here and can be downloaded as follows:

wget https://sharing.speed.pub.ro/owncloud/index.php/s/MTSTY7tnKqn21h4/download -O dolos-models-patches.zip
unzip dolos-models-patches.zip

The name comes from Greek mythology; according to Wikipedia:

Dolos (Ancient Greek: Δόλος "Deception") is the spirit of trickery. He is also a master at cunning deception, craftiness, and treachery. [...] Dolos became known for his skill when he attempted to make a fraudulent copy statue of Aletheia (Veritas), in order to trick people into thinking they were seeing the real statue.

The code and dataset are licensed under CC BY-NC-SA 4.0

The dataset is built on CelebA-HQ and FFHQ datasets introduced in these papers:

Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen. Progressive Growing of GANs for Improved Quality, Stability, and Variation. ICLR, 2018

Tero Karras, Samuli Laine, Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. CVPR, 2019