Self-Supervised Learning for Detecting AI-Generated Faces as Anomalies

This repository contains the official PyTorch implementation of the paper "Self-Supervised Learning for Detecting AI-Generated Faces as Anomalies" by Mian Zou, Baosheng Yu, Yibing Zhan, and Kede Ma.

☀️ If you find this work useful for your research, please kindly star our repo and cite our paper! 😊

⭐️Our series work: BLADES (ICCV'25 Highlight), SDAIE (TPAMI'26)

TODO

We are working hard on the following items.

• Release arXiv paper
• Release training codes
• Release inference codes
• Release checkpoints
• Release datasets

Introduction

In this paper, we describe an anomaly detection method for AI-generated faces by leveraging self-supervised learning of camera-intrinsic and face-specific features purely from photographic face images. The success of our method lies in designing a pretext task that trains a feature extractor to rank four ordinal exchangeable image file format (EXIF) tags and classify artificially manipulated face images. Subsequently, we model the learned feature distribution of photographic face images using a Gaussian mixture model. Faces with low likelihoods are flagged as AI-generated.

📁 Datasets

Dataset & Material	Link
FDF	GitHub	✅
FDF_Artifical_Manipulations	Baidu Disk	✅
Test Sets	Baidu Disk	✅
fdf_ccby2_exif_update_filtered_v2	Baidu Disk	✅
id_vs_fdfName.pkl	Baidu Disk	✅

We use FDF (CC BY-NC-SA 2.0 version) for self-supervised training, and you can download it from the original project page. We also provide the augmented FDF by the proposed artificial face manipulations. After downloading all the necessary files, please put them into the SSL_training_data folder, with the data structure as follows:

SSL_training_data
├── fdf_ccby2_exif_update_filtered_v2
│   ├──10000004756.json
│   ├──...
├── FDF
│   ├── data
│   │   ├──id_vs_fdfName.pkl
│   │   ├── face_dataaug_neg
│   │   │   ├── 1000002_0.png
│   │   │   ├──...
│   │   ├── fdf
│   │   │   ├── bounding_box
│   │   │   ├── landmarks
│   │   │   ├── fdf_metainfo.json
│   │   │   ├── images
│   │   │   │   ├── 128
│   │   │   │   │   ├──100000.png
│   │   │   │   │   ├──...
│   │   │   │   ├──...

For the test sets we used in the main experiments, we collected them from DiffusionFace and DiFF, and put them together for testing. If you find them useful, please cite these two papers. During testing, please put the dataset in the data folder, and the data structure is as follows:

data
├── celeba-face
├── DDIM
├── FreeDoM_T
├── HPS
├── LDM
├── Midjourney
├── SDXL
├── stable_diffusion_v_2_1_text2img_p2g3
├── stylegan2
├── VQ-GAN_celebahq
├── train.txt
├── test.txt

🚀 Quick Start

1. Installation of base reqiurements

• python == 3.8
• PyTorch == 1.13
• Miniconda
• CUDA == 11.7

2. Download the pretrained model and our model

Model	Download
EXIF-pretrained	Google Drive
our model	Google Drive

After downloading these checkpoints, put them into the folder pretrained

3. Inference on the test sets

CUDA_VISIBEL_DEVICES=XXX python yfcc_face_OCT_variants.py --resume [path to our model checkpoints]

⚡ Self-Supervised Training

CUDA_VISIBLE_DEVICES=XXX python train_fidelityCLS.py --batch_size 256 --lr 1e-5
--output [path to save the training files, e.g., checkpoints, training log]
--name [project_name] 

📜 License

This project is released for academic and research purposes only, and is released under the CC BY-NC-SA 2.0 license.

Citation

If you find this repository useful in your research, please consider citing the following paper:

@article{zou2025self,
  title={Self-Supervised Learning for Detecting AI-Generated Faces as Anomalies},
  author={Zou, Mian and Yu, Baosheng and Zhan, Yibing and Ma, Kede},
  journal={arXiv preprint arXiv:2501.02207},
  year={2025}
}