Mega-MMDF and DeepfakeBench-MM: Benchmarking Multimodal Deepfake Detection Across Diverse Forgery Pipelines

🏠 Project homepage  ·  💻 Benchmark  ·  📄 Paper  ·  ⚖️ License

This is the official dataset repository for Mega-MMDF, a million-scale audio-visual deepfake dataset for multimodal deepfake detection research. The accompanying DeepfakeBench-MM benchmark code for preprocessing, training, and evaluation is maintained in the GitHub repository.

🔥 News

• 2026/11/13: We release the Mega-MMDF dataset! 🎉🎉🎉

🔎 Dataset overview

Mega-MMDF is a million-scale audio-visual deepfake dataset with diverse forgery methods and pipelines. It contains 100,000 real and 1.2 million forged clips, constructed using 30 generation methods spanning eight forgery types and organized into 22 forgery pipelines (P1–P22). P0 (pipeline0) corresponds to real data.

To respect the source datasets' requirements, we exclude real data and distribute the fake data from all other pipelines. The release includes sample annotations, source mappings, official train/validation/test splits and a script to restore real data locally from separately obtained source datasets.

🔑 1. Apply for access

Access is limited to individually approved, non-commercial academic research under the License. Each researcher accessing or using the data must apply individually, and requests are reviewed manually. Do not redistribute restricted data or share credentials. To request access to Mega-MMDF, you must complete both application steps below:

1. Sign in to your personal Hugging Face account and submit an access request on this dataset page.
2. Complete and sign the access agreement. Send it from the contact email listed on your form to kangranzhao@link.cuhk.edu.cn with the subject Mega-MMDF access / YOUR_USERNAME.

📥 2. Download and extract

The fake data is packaged as split ZIP archives. FAFV, FARV and RAFV can be downloaded independently; all volumes of a selected category are required.

After approval, sign in to the Hugging Face CLI with your approved account, then download:

hf download SCLBD/Mega-MMDF --repo-type dataset --local-dir Mega-MMDF

Use 7-Zip to extract the archives. Keep all volumes of each category in the same folder, open the first volume (e.g., FAFV/Mega-MMDF_FAFV.zip.001), and extract to the downloaded Mega-MMDF folder. 7-Zip reads the remaining volumes automatically.

Keep JSON files in metadata/ and restore_real_data.py at the dataset root.

🔄 3. Restore real data

Obtain the corresponding clips from VoxCeleb2, MEAD, and CelebV-HQ through their providers or authorized channels, following their terms. Mega-MMDF approval does not grant access to these source datasets.

All three datasets must share one parent directory, with the exact folder names below and their original internal paths:

/path/to/real/
├── VoxCeleb2/
├── MEAD/
└── CelebV-HQ/

Keep restore_real_data.py at the dataset root and the source mappings in metadata/attributes.json. Run from your local dataset directory:

python restore_real_data.py /path/to/real

Replace /path/to/real with your source parent directory. Output goes to RARV/pipeline0/<group>/<filename>.mp4. Every run starts over and overwrites existing targets; a missing source file stops the script.

🗃️ 4. Metadata and splits

• metadata/attributes.json: per-sample metadata, including generation methods and raw_data source mappings for real data.
• metadata/pair.json: 50,000 source-clip pairs, with paths relative to RARV/pipeline0/.
• metadata/train.json, metadata/val.json, metadata/test.json: official splits, including real data. Restore real data to use the complete splits.

Split	Fake samples	Real samples	Total
Train	720,866	60,000	780,866
Validation	239,163	20,000	259,163
Test	239,971	20,000	259,971
Total	1,200,000	100,000	1,300,000

Resolve the Mega-MMDF/ prefix in metadata paths to your local dataset root. Report the dataset version and any subset used in your experiments.

✉️ Updates and contact

Mega-MMDF is developed by the Secure Computing Lab of Big Data (SCLBD) at The Chinese University of Hong Kong, Shenzhen, led by Professor Baoyuan Wu. The laboratory studies trustworthy AI, including adversarial machine learning, agent security and safety, deepfake and AIGC detection, and data governance.

Keep your contact email current and follow emailed update and removal instructions. Material License updates normally allow at least 30 calendar days for explicit acceptance. Refusal or failure to confirm leads to termination as notified; stop use when termination takes effect and complete cleanup, normally within 14 calendar days, including covered older releases and archive copies. Continuing obligations and responsibility for violations remain in effect. See License Sections 5-6 for details.

For access questions, contact kangranzhao@link.cuhk.edu.cn. For privacy or removal requests, use the same address and copy wubaoyuan@cuhk.edu.cn, with enough information to locate the material and explain the request. No dataset access is needed to submit a request. Approved removals are followed by an updated release, user notification, and an update script.

📝 Citation

If you use Mega-MMDF or DeepfakeBench-MM in your research, please cite the following work:

@misc{megammdf,
  title={Mega-MMDF and DeepfakeBench-MM: Benchmarking Multimodal Deepfake Detection Across Diverse Forgery Pipelines},
  author={Zhao, Kangran and
          Zhang, Xiaoyu and
          Chen, Yupeng and
          Chen, Yize and
          Guan, Weinan and
          Zheng, Bo and
          Liao, Yufan and
          Sun, Chengzhe and
          Datta, Soumyya Kanti and
          Liu, Qingshan and
          Lyu, Siwei and
          Wu, Baoyuan},
  note={Manuscript; publication details to be updated}
}

🗓️ Release history

Version	Release date	Distributed fake samples
v1.0	2026/11/13	1,200,000