Skip to content

Latest commit

 

History

History
197 lines (152 loc) · 21.6 KB

README.md

File metadata and controls

197 lines (152 loc) · 21.6 KB

ArtiFact: A Large-Scale Dataset with Artificial and Factual Images for Generalizable and Robust Synthetic Image Detection [ICIP 2023]

Paper:

Abstract: Synthetic image generation has opened up new opportunities but has also created threats in regard to privacy, authenticity, and security. Detecting fake images is of paramount importance to prevent illegal activities, and previous research has shown that generative models leave unique patterns in their synthetic images that can be exploited to detect them. However, the fundamental problem of generalization remains, as even state-of-the-art detectors encounter difficulty when facing generators never seen during training. To assess the generalizability and robustness of synthetic image detectors in the face of real-world impairments, this paper presents a large-scale dataset named ArtiFact, comprising diverse generators, object categories, and real-world challenges. Moreover, the proposed multi-class classification scheme, combined with a filter stride reduction strategy addresses social platform impairments and effectively detects synthetic images from both seen and unseen generators. The proposed solution significantly outperforms other top teams by 8.34% on Test 1, 1.26% on Test 2, and 15.08% on Test 3 in the IEEE VIP Cup challenge at ICIP 2022, as measured by the accuracy metric.

Presentation: YouTube

Visual Summary:

Update

  • [22 June 2023] - The work has been accepted to IEEE ICIP 2023 conference.

Accuracy (%) of Top3 Teams on Leaderboard,

Team Names Test 1 Test 2 Test 3
Sherlock 87.70 77.52 73.45
FAU Erlangen-Nürnberg 87.14 81.74 75.52
Megatron (Ours) 96.04 83.00 90.60

Note: A small portion of the proposed ArtiFact dataset, totaling 222K images of 71K real images and 151K fake images from only 13 generators is used in the IEEE VIP Cup. Here all the Test data is kept confidential from all participating teams. Additionally, the generators used for the Test 1 data are known to all teams, whereas the generators for Test 2 and Test 3 are kept undisclosed.

Dataset Description

  • Total number of images: $2,496,738$
  • Number of real images: $964,989$
  • Number of fake images: $1,531,749$
  • Number of generators used for fake images: $25$ (including $13$ GANs, $7$ Diffusion, and $5$ miscellaneous generators)
  • Number of sources used for real images: $8$
  • Categories included in the dataset: Human/Human Faces, Animal/Animal Faces, Places, Vehicles, Art, and other real-life objects
  • Image Resolution: $200 \times 200$

Data Distribution

  • Real
  • Fake

Download Dataset

The dataset is hosted on Kaggle. The dataset can be downloaded i) directly from the browser using the link below or ii) can be downloaded using kaggle-api.

i) Directly from Browser

Link: ArtiFact Dataset

ii) Kaggle API

!kaggle datasets download -d awsaf49/artifact-dataset

How to Use

The dataset is organized into folders, each of which corresponds to a specific generator of synthetic images or source of real images. Each folder contains a metadata.csv file, which provides information about the images in the folder. It contains following columns,

  • image_path : The relative path of the image file.
  • target : The label for the image, which is either 0 for real or 1 for fake.
  • category : The category (cat or dog etc) of the image

Data Generation

  • Images are randomly sampled from different methods then transformed using impairments. The methods are listed below,

    Methods
    Method ImageNet COCO LSUN AFHQ FFHQ Metfaces CelebAHQ Landscape Glide StyleGAN2 StyleGAN3 Generative Inpainting Taming Transformer MAT LaMa Stable Diffusion VQ Diffusion Palette StyleGAN1 Latent Diffusion CIPS StarGAN BigGAN GANformer ProjectedGAN SFHQ FaceSynthetics Denoising Diffusion GAN DDPM DiffusionGAN GauGAN ProGAN CycleGAN
    Reference link link link link link link link link link link link link link link link link link link link link link link link link link link link link link link link link link
  • All images went through RandomCrop and Random Impairments (Jpeg Compression & Downscale). To apply these transformation use data/transform.py which applies random transformation. All images are cropped and resized to $200 \times 200$ pixels and then compressed using JPEG at a random quality level.

!python data/transform.py <input directory> <output directory> <seed>

Citation

@INPROCEEDINGS{artifact,
  author={Rahman, Md Awsafur and Paul, Bishmoy and Sarker, Najibul Haque and Hakim, Zaber Ibn Abdul and Fattah, Shaikh Anowarul},
  booktitle={2023 IEEE International Conference on Image Processing (ICIP)}, 
  title={Artifact: A Large-Scale Dataset With Artificial And Factual Images For Generalizable And Robust Synthetic Image Detection}, 
  year={2023},
  volume={},
  number={},
  pages={2200-2204},
  doi={10.1109/ICIP49359.2023.10222083}}

License

ArtiFact dataset takes leverage of data from multiple methods thus different parts of the dataset come with different licenses. All the methods and their associated licenses are mentioned in the table,

Data License
Method License
ImageNet Non Commercial
COCO Creative Commons Attribution 4.0 License
LSUN Unknown
AFHQ Creative Commons Attribution-NonCommercial 4.0 International Public
FFHQ Creative Commons BY-NC-SA 4.0 license
Metfaces Creative Commons BY-NC 2.0
CelebAHQ Creative Commons Attribution-NonCommercial 4.0 International Public
Landscape MIT license
Glide MIT license
StyleGAN2 Nvidia Source Code License
StyleGAN3 Nvidia Source Code License
Generative Inpainting Creative Commons Public Licenses
Taming Transformer MIT License
MAT Creative Commons Public Licenses
LaMa Apache-2.0 License
Stable Diffusion Apache-2.0 License
VQ Diffusion MIT License
Palette MIT License
StyleGAN1 Creative Commons Public Licenses
Latent Diffusion MIT License
CIPS MIT License
StarGAN MIT License
BigGAN MIT License
GANformer MIT License
ProjectedGAN MIT License
SFHQ MIT License
FaceSynthetics Research Use of Data Agreement v1.0
Denoising Diffusion GAN NVIDIA License
DDPM Unknown
DiffusionGAN MIT License
GauGAN Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License
ProGAN Attribution-NonCommercial 4.0 International
CycleGAN BSD

Acknowledgment

  • The authors would like to express their gratitude to the IEEE Signal Processing Society, GRIP of the University Federico II of Naples (Italy), and NVIDIA (USA) for hosting the IEEE Video and Image Processing (VIP) Cup competition at ICIP 2022. This competition provided a platform for the authors to showcase their work and motivated them to push their boundaries to deliver a state-of-the-art solution.

  • The authors also would like to express their gratitude to the authors of the methods that is used for creating ArtiFact dataset. All the methods and their reference is added below,

    Data Reference
    Method Reference
    ImageNet link
    COCO link
    LSUN link
    AFHQ link
    FFHQ link
    Metfaces link
    CelebAHQ link
    Landscape link
    Glide link
    StyleGAN2 link
    StyleGAN3 link
    Generative Inpainting link
    Taming Transformer link
    MAT link
    LaMa link
    Stable Diffusion link
    VQ Diffusion link
    Palette link
    StyleGAN1 link
    Latent Diffusion link
    CIPS link
    StarGAN link
    BigGAN link
    GANformer link
    ProjectedGAN link
    SFHQ link
    FaceSynthetics link
    Denoising Diffusion GAN link
    DDPM link
    DiffusionGAN link
    GauGAN link
    ProGAN link
    CycleGAN link