AADD-2025: Adversarial Attacks on Deepfake Detectors Challenge

1st Adversarial Attacks on Deepfake Detectors: A Challenge in the Era of AI-Generated Media

Grand Challenge at ACM Multimedia 2025

---

🎯 Overview

The AADD-2025 Challenge investigated adversarial vulnerabilities of deepfake detection models by generating adversarial perturbed deepfake images that evade standard classifiers while maintaining high visual similarity to the original deepfake content. Given the increasing reliance on deepfake detectors in forensic analysis and content moderation, ensuring their robustness against adversarial attacks has relevant importance.

🎪 Challenge Description

The goal of this challenge was to expose and address vulnerabilities in current deepfake detection systems by designing adversarial attacks that alter deepfake images, rendering them unrecognizable as synthetic content to 4 proposed classifiers, preserving high visual similarity to the original images.

📊 Dataset Structure

Participants were provided with a dataset divided into sixteen subsets:

High Quality Resolution:

• 4 GAN-based models (high quality)
• 4 Diffusion-based models (high quality)

Low Quality Resolution:

• 4 GAN-based models (low quality)
• 4 Diffusion-based models (low quality)

- Dataset
├── train
│   ├── fake
│   │   ├── hq
│   │   │   ├── Adobe Firefly
│   │   │   ├── Deep AI
│   │   │   ├── Flux.1.1 Pro
│   │   │   ├── Hotpot AI
│   │   │   ├── Nvidia Sana PAG
│   │   │   ├── Stable Diffusion 3.5
│   │   │   ├── StyleGAN2
│   │   │   ├── StyleGAN3
│   │   │   └── Tencent Hunyuan
│   │   └── lq
│   │       ├── Deep AI
│   │       ├── Flux.1
│   │       ├── Freepik
│   │       ├── Hotpot AI
│   │       ├── Nvidia Sana PAG
│   │       ├── Stable Diffusion Attend and Excite
│   │       ├── StyleGAN
│   │       ├── StyleGAN3
│   │       └── Tencent Hunyuan
│   └── real
│       ├── hq
│       │   └── ffhq
│       └── lq
│           └── celeba_hq
└── test
    ├── fake
    │   ├── hq
    │   │   ├── Adobe Firefly
    │   │   ├── Deep AI
    │   │   ├── Flux.1.1 Pro
    │   │   ├── Hotpot AI
    │   │   ├── Nvidia Sana PAG
    │   │   ├── Stable Diffusion 3.5
    │   │   ├── StyleGAN2
    │   │   ├── StyleGAN3
    │   │   └── Tencent Hunyuan
    │   └── lq
    │       ├── Deep AI
    │       ├── Flux.1
    │       ├── Freepik
    │       ├── Hotpot AI
    │       ├── Nvidia Sana PAG
    │       ├── Stable Diffusion Attend and Excite
    │       ├── StyleGAN
    │       ├── StyleGAN3
    │       └── Tencent Hunyuan
    └── real
        ├── hq
        │   └── ffhq
        └── lq
            └── celeba_hq

Note: Participants had to focus on the entire dataset across all subsets.

📋 Submission Requirements

1. Adversarial Images: Submit the generated adversarial deepfake images
2. Technical Abstract: Provide a detailed description of your methodology
3. Results Documentation: Include performance metrics and analysis

📥 Evaluation Resources

Final Evaluation Scripts See here

🏆 Results & Rankings

The challenge ended with strong global participation. Here are the final standings:

Rank	Team Name	Organization/Institution	Final Score
🥇 1st	MR-CAS	🇨🇳 University of Chinese Academy of Sciences	2740
🥈 2nd	Safe AI	🇰🇷 UNIST (Ulsan National Institute of Science and Technology)	2709
🥉 3rd	RoMa	🇩🇪 Fraunhofer SIT | ATHENE Center	2679
4th	GRADIANT	🇪🇸 Gradiant	2631
5th	DASH	🇰🇷 Sungkyunkwan University	2618
6th	SecureML	🇮🇹 University of Cagliari	2490
7th	MICV	🇨🇳 Ant Group	2434
8th	WHU_PB	🇨🇳 Wuhan University	2354
9th	The Adversaries	🇸🇬 Singapore Institute of Technology	2341
10th	DeFakePol	🇵🇱 Samsung Research Poland	1665
11th	False Negative	🇨🇳 The Hong Kong Polytechnic University	1602
12th	VYAKRITI 2.0	🇮🇳 Apex Institute of technology Chandigarh University	1041
13th	MILab	🇨🇳 University of Science and Technology of China	110

📊 Timeline

The AADD-2025 Challenge followed this timeline:

• ✅ March 03, 2025: Competition Website Launch
• ✅ March 15 - May 22, 2025: Registration Period (Extended)
• ✅ April 10, 2025: Test Set and Classificator Release
• ✅ June 15, 2025: Final Submission Deadline
• ✅ June 22, 2025: Leaderboard Publication and Rankings Release
• ✅ June 30, 2025: Paper Submission Deadline (Top 3 Teams Only)
• ⏳ July 24, 2025: Announcement regarding full paper submission
• ⏳ August 01, 2025: Camera ready - Grand Challenge Solutions (Top 3 Teams Only)
• ⏳ ACM Multimedia 2025: Conference & Winners Recognition

📝 Publication Opportunities

The top 3 teams were invited to submit full-length papers describing their methods in detail. These papers underwent a rigorous review process managed by the challenge organizers, with accepted papers included in the ACM Multimedia 2025 proceedings.

👥 Organizing Committee

Chairs

Name	Role	Email	Affiliation
Luca Guarnera	Research Fellow	[email protected]	Department of Mathematics and Computer Science, University of Catania, Italy
Francesco Guarnera	Research Fellow	[email protected]	Department of Mathematics and Computer Science, University of Catania, Italy

Co-Chairs

Name	Role	Email	Affiliation
Sebastiano Battiato	Full Professor	[email protected]	Department of Mathematics and Computer Science, University of Catania, Italy
Giovanni Puglisi	Associate Professor	[email protected]	Department of Mathematics and Informatics, University of Cagliari, Italy
Zahid Akhtar	Associate Professor	[email protected]	State University of New York Polytechnic Institute, USA

Technical Committee

Name	Role	Email	Affiliation
Mirko Casu	PhD Student	[email protected]	Department of Mathematics and Computer Science, University of Catania, Italy
Orazio Pontorno	PhD Student	[email protected]	Department of Mathematics and Computer Science, University of Catania, Italy
Claudio Vittorio Ragaglia	PhD Student	[email protected]	Department of Mathematics and Computer Science, University of Catania, Italy

📧 Contact Information

Main Contact: Mirko Casu
Email: [email protected]

📖 Citation

Dataset Attribution: Part of this challenge dataset is based on the WILD dataset. If you use the data, please also cite:

@misc{bongini2025wildnewinthewildimage,
      title={WILD: a new in-the-Wild Image Linkage Dataset for synthetic image attribution}, 
      author={Pietro Bongini and Sara Mandelli and Andrea Montibeller and Mirko Casu and Orazio Pontorno and Claudio Vittorio Ragaglia and Luca Zanchetta and Mattia Aquilina and Taiba Majid Wani and Luca Guarnera and Benedetta Tondi and Giulia Boato and Paolo Bestagini and Irene Amerini and Francesco De Natale and Sebastiano Battiato and Mauro Barni},
      year={2025},
      eprint={2504.19595},
      archivePrefix={arXiv},
      primaryClass={cs.MM},
      url={https://arxiv.org/abs/2504.19595}, 
}

If you use or refer to this challenge, please cite our ACM Multimedia 2025 paper:

@inproceedings{battiato2025adversarial,
  title={Adversarial Attacks on Deepfake Detectors: A Challenge in the Era of AI-Generated Media (AADD-2025)},
  author={Battiato, Sebastiano and Casu, Mirko and Guarnera, Francesco and Guarnera, Luca and Puglisi, Giovanni and Pontorno, Orazio and Ragaglia, Claudio Vittorio and Akhtar, Zahid},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={13714--13719},
  year={2025}
}

🌐 Related Resources

• ACM Multimedia 2025 Conference
• Challenge Website

Institutional Affiliations:

• University of Catania - Department of Mathematics and Computer Science
• University of Cagliari - Department of Mathematics and Informatics
• State University of New York Polytechnic Institute - Website

© 2025 University of Catania.
Powered by the Multimedia Security and Forensics group of the Image Processing Laboratory (IPLAB).

🏷️ Keywords

Deepfake Detection, Adversarial Attacks, Computer Vision, Digital Forensics, AI Security, Media Authentication, Challenge Competition, ACM Multimedia

---