SAFE: Simple Preserved and Augmented FEatures

This is the official Pytorch implementation of our paper:

Improving Synthetic Image Detection Towards Generalization: An Image Transformation Perspectives

Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, Fuli Feng

News

• 2024/11 🔥 We collect a new testset DiTFake, comprising three SOTA DiT-based generators (i.e., Flux, PixArt, and SD3). We hope this dataset could facilitate more comprehensive evaluations for SID.
• 2024/11 🎉 Our paper is accepted by KDD2025 ADS Track.

Requirements

Install the environment as follows:

# create conda environment
conda create -n SAFE -y python=3.9
conda activate SAFE
# install pytorch 
pip install torch==2.2.1 torchvision==0.17.1
# install other dependencies
pip install -r requirements.txt

We are using torch 2.2.1 in our production environment, but other versions should be fine as well.

Getting the data

	paper	Url
Train Set	CNNDetection CVPR2020	Link
Val Set	CNNDetection CVPR2020	Link
Test Set1	CNNDetection CVPR2020	Link
Test Set2	FreqNet AAAI2024	Link
Test Set3	UniversalFakeDetect CVPR2023	Link
Test Set4	GenImage NeurIPS2023	Link
Test Set5	DiTFake Ours	Link

The generation script for our dataset is provided in data/generation.py, we hope more synthetic images from up-to-date generative models coud be promptly evaluated and made publicly available. Details of our DiTFake testset and comparative results will be presented in the forthcoming camera-ready version soon.

Directory structure

You should organize the above data as follows:

data/datasets
|-- train_ForenSynths
|   |-- train
|   |   |-- car
|   |   |-- cat
|   |   |-- chair
|   |   |-- horse
|   |-- val
|   |   |-- car
|   |   |-- cat
|   |   |-- chair
|   |   |-- horse
|-- test1_ForenSynths/test
|   |-- biggan
|   |-- cyclegan
|   |-- deepfake
|   |-- gaugan
|   |-- progan
|   |-- stargan
|   |-- stylegan
|   |-- stylegan2
|-- test2_Self-Synthesis/test
|   |-- AttGAN
|   |-- BEGAN
|   |-- CramerGAN
|   |-- InfoMaxGAN
|   |-- MMDGAN
|   |-- RelGAN
|   |-- S3GAN
|   |-- SNGAN
|   |-- STGAN
|-- test3_Ojha/test
|   |-- dalle
|   |-- glide_100_10
|   |-- glide_100_27
|   |-- glide_50_27
|   |-- guided          # Also known as ADM.
|   |-- ldm_100
|   |-- ldm_200
|   |-- ldm_200_cfg
|-- test4_GenImage/test
|   |-- ADM
|   |-- BigGAN
|   |-- Glide
|   |-- Midjourney
|   |-- stable_diffusion_v_1_4
|   |-- stable_diffusion_v_1_5
|   |-- VQDM
|   |-- wukong
|-- test5_DiTFake/test
|   |-- FLUX.1-schnell
|   |-- PixArt-Sigma-XL-2-1024-MS
|   |-- stable-diffusion-3-medium-diffusers

Training

bash scripts/train.sh

This script enables training with 4 GPUs, you can specify the number of GPUs by setting GPU_NUM.

Inference

bash scripts/eval.sh

We provide the pretrained checkpoint in ./checkpoint/checkpoint-best.pth, you can directly run the script to reproduce our results.

Citing

If you find this repository useful for your work, please consider citing it as follows:

@article{li2024improving,
  title={Improving Synthetic Image Detection Towards Generalization: An Image Transformation Perspective},
  author={Li, Ouxiang and Cai, Jiayin and Hao, Yanbin and Jiang, Xiaolong and Hu, Yao and Feng, Fuli},
  journal={arXiv preprint arXiv:2408.06741},
  year={2024}
}