AV-Deepfake1M

This is the official repository for the paper AV-Deepfake1M: A Large-Scale LLM-Driven Audio-Visual Deepfake Dataset (Best Award).

News

• [2026/03/22] 🔥 2026 Deepfake Explainability Challenge starts in MM 2026.
• [2025/08/03] 🔥 AV-Deepfake1M++ technical report is in arXiv.
• [2025/04/08] 🏆 2025 1M-Deepfakes Challenge starts. New dataset (2M videos!) - AV-Deepfake1M++ is released.
• [2024/10/13] 🚀 PYPI package is released.
• [2024/07/15] 🔥 AV-Deepfake1M paper is accepted in MM 2024.
• [2024/03/09] 🏆 2024 1M-Deepfakes Challenge starts.

Abstract

The detection and localization of highly realistic deepfake audio-visual content are challenging even for the most advanced state-of-the-art methods. While most of the research efforts in this domain are focused on detecting high-quality deepfake images and videos, only a few works address the problem of the localization of small segments of audio-visual manipulations embedded in real videos. In this research, we emulate the process of such content generation and propose the AV-Deepfake1M dataset. The dataset contains content-driven (i) video manipulations, (ii) audio manipulations, and (iii) audio-visual manipulations for more than 2K subjects resulting in a total of more than 1M videos. The paper provides a thorough description of the proposed data generation pipeline accompanied by a rigorous analysis of the quality of the generated data. The comprehensive benchmark of the proposed dataset utilizing state-of-the-art deepfake detection and localization methods indicates a significant drop in performance compared to previous datasets. The proposed dataset will play a vital role in building the next-generation deepfake localization methods.

demo.mp4

Dataset

Download

We're hosting 1M-Deepfakes Detection Challenge at ACM MM 2024.

Baseline Benchmark

Method	AP@0.5	AP@0.75	AP@0.9	AP@0.95	AR@50	AR@20	AR@10	AR@5
PyAnnote	00.03	00.00	00.00	00.00	00.67	00.67	00.67	00.67
Meso4	09.86	06.05	02.22	00.59	38.92	38.81	36.47	26.91
MesoInception4	08.50	05.16	01.89	00.50	39.27	39.00	35.78	24.59
EfficientViT	14.71	02.42	00.13	00.01	27.04	26.43	23.90	20.31
TriDet + VideoMAEv2	21.67	05.83	00.54	00.06	20.27	20.12	19.50	18.18
TriDet + InternVideo	29.66	09.02	00.79	00.09	24.08	23.96	23.50	22.55
ActionFormer + VideoMAEv2	20.24	05.73	00.57	00.07	19.97	19.81	19.11	17.80
ActionFormer + InternVideo	36.08	12.01	01.23	00.16	27.11	27.00	26.60	25.80
BA-TFD	37.37	06.34	00.19	00.02	45.55	35.95	30.66	26.82
BA-TFD+	44.42	13.64	00.48	00.03	48.86	40.37	34.67	29.88
UMMAFormer	51.64	28.07	07.65	01.58	44.07	43.45	42.09	40.27

Metadata Structure

The metadata is a json file for each subset (train, val), which is a list of dictionaries. The fields in the dictionary are as follows.

• file: the path to the video file.
• original: if the current video is fake, the path to the original video; otherwise, the original path in VoxCeleb2.
• split: the name of the current subset.
• modify_type: the type of modifications in different modalities, which can be ["real", "visual_modified", "audio_modified", "both_modified"]. We evaluate the deepfake detection performance based on this field.
• audio_model: the audio generation model used for generating this video.
• fake_segments: the timestamps of the fake segments. We evaluate the temporal localization performance based on this field.
• audio_fake_segments: the timestamps of the fake segments in audio modality.
• visual_fake_segments: the timestamps of the fake segments in visual modality.
• video_frames: the number of frames in the video.
• audio_frames: the number of frames in the audio.

SDK

We provide a Python library avdeepfake1m to load the dataset and evaluation.

Installation

pip install avdeepfake1m

Usage

Prepare the dataset as follows.

|- train_metadata.json
|- train_metadata
| |- ...
|- train
| |- ...
|- val_metadata.json
|- val_metadata
| |- ...
|- val
| |- ...
|- test_files.txt
|- test

Load the dataset.

fromavdeepfake1m.loaderimportAVDeepfake1mDataModule# access to Lightning DataModuledm=AVDeepfake1mDataModule("/path/to/dataset")

Evaluate the predictions. Firstly prepare the predictions as described in the details. Then run the following code.

fromavdeepfake1m.evaluationimportap_ar_1d, aucprint(ap_ar_1d("<PREDICTION_JSON>", "<METADATA_JSON>", "file", "fake_segments", 1, [0.5, 0.75, 0.9, 0.95], [50, 30, 20, 10, 5], [0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95]))
print(auc("<PREDICTION_TXT>", "<METADATA_JSON>", "file", "fake_segments"))

License

The dataset is under the EULA. You need to agree and sign the EULA to access the dataset.

The baseline Xception code /examples/xception is under MIT Licence. The BA-TFD/BA-TFD+ code /examples/batfd is from ControlNet/LAV-DF under CC BY-NC 4.0 Licence.

The other parts of this project is under the CC BY-NC 4.0 license. See LICENSE for details.

References

If you find this work useful in your research, please cite it.

The AV-Deepfake1M++ dataset paper:

@article{cai2025av,
title={AV-Deepfake1M++: A Large-Scale Audio-Visual Deepfake Benchmark with Real-World Perturbations},
author={Cai, Zhixi and Kuckreja, Kartik and Ghosh, Shreya and Chuchra, Akanksha and Khan, Muhammad Haris and Tariq, Usman and Gedeon, Tom and Dhall, Abhinav},
journal={arXiv preprint arXiv:2507.20579},
year={2025}
}

The AV-Deepfake1M dataset paper:

@inproceedings{cai2024av,
title={AV-Deepfake1M: A large-scale LLM-driven audio-visual deepfake dataset},
author={Cai, Zhixi and Ghosh, Shreya and Adatia, Aman Pankaj and Hayat, Munawar and Dhall, Abhinav and Gedeon, Tom and Stefanov, Kalin},
booktitle={Proceedings of the 32nd ACM International Conference on Multimedia},
pages={7414--7423},
year={2024},
doi={10.1145/3664647.3680795}
}

The challenge summary paper:

@inproceedings{cai20241m,
title={1M-Deepfakes Detection Challenge},
author={Cai, Zhixi and Dhall, Abhinav and Ghosh, Shreya and Hayat, Munawar and Kollias, Dimitrios and Stefanov, Kalin and Tariq, Usman},
booktitle={Proceedings of the 32nd ACM International Conference on Multimedia},
pages={11355--11359},
year={2024},
doi={10.1145/3664647.3689145}
}