Diffusion Facial Forgery Detection

DiFF — a benchmark of 500,000+ diffusion-generated facial forgeries across 13 methods and 4 conditions, exposing the challenge of detecting face-focused diffusion fakes.

Figure 1. DiFF dataset overview.

Authors

Harry Cheng¹, Yangyang Guo²*, Tianyi Wang³, Liqiang Nie⁴*, Mohan Kankanhalli²

¹ Shandong University ² National University of Singapore ³ Nanyang Technological University ⁴ Harbin Institute of Technology (Shenzhen) * Corresponding authors

Links

• Paper: Diffusion Facial Forgery Detection
• arXiv: 2401.15859
• Dataset (synthesized images): Google Drive
• Dataset (pristine images & prompts): Request Access
• Code Repository: GitHub

---

Table of Contents

• Updates
• Introduction
• Highlights
• Method / Framework
• Dataset / Benchmark
• Citation
• Acknowledgement
• License

---

Updates

• [04/2026] Transfer repos to iLearn-Lab
• [2024] Paper accepted at ACM MM 2024

---

Introduction

This repository presents DiFF (Diffusion Facial Forgery Detection), accepted at ACM MM 2024.

Detecting diffusion-generated images has emerged as a critical research area, yet existing benchmarks focus on general image generation. Face-focused forgeries — which pose a more severe social risk — have remained underexplored. DiFF fills this gap with a comprehensive benchmark dedicated to diffusion-generated facial images.

DiFF comprises over 500,000 images synthesized by 13 distinct generation methods under 4 conditions, using 30,000 carefully collected textual and visual prompts to ensure semantic consistency and high fidelity.

Our experiments show that binary detection accuracy for both human observers and automated detectors often falls below 30%, highlighting the difficulty of this problem. We also propose an edge graph regularization approach to improve detector generalization.

---

Highlights

• First large-scale benchmark for face-focused diffusion-generated forgery detection
• 500,000+ images from 13 generation methods under 4 conditions
• Human and automated detectors both achieve accuracy below 30% on binary detection
• Proposes edge graph regularization to improve detector generalization

---

Method / Framework

Visual Examples

Figure 2. Example images from DiFF across different methods and conditions.

---

Dataset / Benchmark

Synthesized Images

Directly downloadable:

• Google Drive: Link

Pristine Images & Prompts

Access via request (to protect privacy):

• Google Form: Request Access

---

Citation

If you find this work helpful, please cite:

@inproceedings{cheng2024diffusion,
  title     = {Diffusion Facial Forgery Detection},
  author    = {Cheng, Harry and Guo, Yangyang and Wang, Tianyi and Nie, Liqiang and Kankanhalli, Mohan},
  booktitle = {Proceedings of the ACM International Conference on Multimedia},
  year      = {2024},
  pages     = {5939--5948},
}

---

Acknowledgement

• Thanks to the authors of the 13 diffusion-based generation methods included in DiFF.
• Thanks to our supervisors and collaborators for their support.

---

License

This project is released under the Apache License 2.0.