Hugging Face's logo

Datasets:
kartik060702
/
ArEnAV-Full

Modalities:
Video
Audio
Languages:
Arabic
English
ArXiv:
Tags:
deepfakes
video
audio
License:
Dataset card Files Community

Access ArEnAV dataset on Hugging Face

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

To use this ArEnAV dataset, you should agree to the CC BY-NC 4.0 license sign the EULA form and email it to the organizers.

Log in or Sign Up to review the conditions and access this dataset content.

Tell me Habibi, is it Real or Fake?

Kartik Kuckreja, Parul Gupta, Injy Hamed, Thamar Solorio, Muhammad Haris Khan and Abhinav Dhall

Mohamed bin Zayed University of AI, Monash University

Abstract

Deepfake generation methods are evolving fast, making fake media harder to detect and raising serious societal concerns. Most deepfake detection and dataset creation research focuses on monolingual content, often overlooking the challenges of multilingual and code-switched speech, where multiple languages are mixed within the same discourse. Code-switching, especially between Arabic and English, is common in the Arab world and is widely used in digital communication. This linguistic mixing poses extra challenges for deepfake detection, as it can confuse models trained mostly on monolingual data. To address this, we introduce ArEnAV, the first large-scale Arabic-English audio-visual deepfake dataset featuring intra-utterance code-switching, dialectal variation, and monolingual Arabic content. It contains 387k videos and over 765 hours of real and fake videos. Our dataset is generated using a novel pipeline integrating four Text-To-Speech and two lip-sync models, enabling comprehensive analysis of multilingual multimodal deepfake detection. We benchmark our dataset against existing monolingual and multilingual datasets, state-of-the-art deepfake detection models, and a human evaluation, highlighting its potential to advance deepfake research.

Cross-Dataset Comparison: Details for publicly available deep - fake datasets in chronologically ascending order.

Cla: Binary classification, SL: Spatial localization, TL: Temporal localization, FS: Face swapping, RE: Face reenactment, TTS: Text-to-speech, VC: Voice conversion.

Dataset Year Tasks Modality Method #Clips Multilingual Code-Switch
Google DFD 2019 Cla V FS 3 431 ✗ ✗
DFDC 2020 Cla AV FS 128 154 ✗ ✗
DeeperForensics 2020 Cla V FS 60 000 ✗ ✗
Celeb-DF 2020 Cla V FS 6 229 ✗ ✗
KoDF 2021 Cla V FS/RE 237 942 ✗ ✗
FakeAVCeleb 2021 Cla AV RE 25 500+ ✗ ✗
ForgeryNet 2021 SL/TL/Cla V FS/RE 221 247 ✗ ✗
ASVSpoof2021-DF 2021 Cla A TTS/VC 593 253 ✗ ✗
LAV-DF 2022 TL/Cla AV RE/TTS 136 304 ✗ ✗
DF-Platter 2023 Cla V FS 265 756 ✗ ✗
AV-1M 2023 TL/Cla AV RE/TTS 1 146 760 ✗ ✗
PolyGlotFake 2024 Cla AV RE/TTS/VC 15 238 ✓ ✗
Illusion 2025 Cla AV FS/RE/TTS 1 376 371 ✓ ✗
ArEnAV (ours) 2025 Cla/TL AV RE/TTS/VC 387 072 ✓ ✓

Multilingual Datasets Comparison: Data distribution in ArEnAV and comparison with other multilingual datasets.

Subset Unique Vids Real Fake Non-Eng CSW Vids Arabic Vids Arabic Variants
PolyGlotFake 766 766 14 472 11 941 0 1 403 –
Illusion – 141 440 1 234 931 4 385 0 – –
Train 6 117 67 600 202 800 270 400 69 544 200 856 Eg, MSA, Lev, Gulf
Val 876 9 560 28 680 38 240 10 416 27 824 Eg, MSA, Lev, Gulf
Test 1 816 19 608 58 824 78 432 19 832 58 600 Eg, MSA, Lev, Gulf
Total 8 809 96 768 290 304 387 072 99 792 287 280 –

Results

Temporal localization results on the test set of ArEnAV.

(Full table preserved for reproducibility)

Set Method Mod. AP @0.5 AP @0.75 AP @0.9 AP @0.95 AR @50 AR @30 AR @20 AR @10 AR @5
Full Meso4 V 0.02 0.01 0.00 0.00 0.09 0.09 0.09 0.09 0.09
MesoInception4 V 0.56 0.18 0.04 0.01 4.11 4.11 4.11 4.11 4.08
Xception V 22.50 10.26 2.29 0.58 19.13 19.13 19.13 19.13 19.13
BA-TFD (ZS) AV 0.17 0.01 0.00 0.00 9.72 5.20 3.07 1.46 0.73
BA-TFD+ (ZS) AV 0.11 0.00 0.00 0.00 5.77 2.95 2.09 0.87 0.37
BA-TFD AV 2.42 0.55 0.01 0.00 22.30 10.31 3.41 2.54 1.67
BA-TFD+ AV 3.74 1.10 0.06 0.01 30.75 9.42 4.55 3.05 1.83
Set V Meso4 V 0.02 0.01 0.00 0.00 0.10 0.10 0.10 0.10 0.10
MesoInception4 V 0.83 0.27 0.05 0.01 5.56 5.56 5.56 5.56 5.53
Xception V 32.76 14.48 3.30 0.81 27.78 27.78 27.78 27.78 27.78
BA-TFD (ZS) AV 0.12 0.00 0.00 0.00 8.44 4.34 2.44 1.13 0.49
BA-TFD+ (ZS) AV 0.07 0.00 0.00 0.00 4.69 2.39 1.65 0.69 0.29
BA-TFD AV 3.65 0.25 0.01 0.00 25.31 9.03 3.64 2.34 1.64
BA-TFD+ AV 5.65 1.89 0.08 0.02 31.09 13.21 5.91 3.05 2.05
Set A Meso4 V 0.02 0.01 0.00 0.00 0.08 0.08 0.08 0.08 0.08
MesoInception4 V 0.38 0.09 0.01 0.00 3.25 3.25 3.25 3.25 3.22
Xception V 14.72 3.92 0.29 0.09 11.78 11.78 11.78 11.78 11.78
BA-TFD (ZS) AV 0.23 0.01 0.00 0.00 12.14 6.46 3.85 1.83 0.95
BA-TFD+ (ZS) AV 0.14 0.01 0.00 0.00 7.32 3.79 2.69 1.13 0.48
BA-TFD AV 3.21 0.60 0.02 0.00 24.45 9.26 4.15 2.61 1.93
BA-TFD+ AV 4.35 1.10 0.10 0.00 28.35 11.23 4.85 3.11 2.00

Deepfake detection results on the test set of ArEnAV.

Access Pre-train Method Mod. Full AUC Full Acc V AUC V Acc A AUC A Acc
Zero-shot ASVSpoof-19 XLSR-Mamba A 39.19 52.77 52.73 40.68 52.50 42.59
Zero-shot Internet Video-LLaMA-7B V 51.48 26.29 51.47 34.21 51.43 34.18
Zero-shot Internet Video-LLaMA-7B AV 48.79 59.29 48.71 55.37 48.86 55.26
Zero-shot AV-1M BA-TFD AV 61.73 26.00 66.42 34.07 59.36 33.97
Zero-shot AV-1M BA-TFD+ AV 60.96 25.84 64.49 34.28 59.44 33.80
Video-level ArEnAV XLSR-Mamba A 73.00 61.00 57.47 66.16 86.33 78.00
Video-level ArEnAV Meso4 V 49.30 75.00 49.15 66.67 49.30 66.67
Video-level ArEnAV MesoInception4 V 50.34 46.23 50.28 47.48 50.35 47.67
Video-level ArEnAV Xception V 50.05 75.00 49.90 66.67 50.32 66.67
Frame-level ArEnAV Meso4 V 49.55 26.60 49.60 34.40 49.53 34.36
Frame-level ArEnAV MesoInception4 V 51.14 41.25 50.77 51.84 45.28 44.09
Frame-level ArEnAV Xception V 74.21 52.09 85.36 67.22 68.59 51.70
Frame-level AV-1M + ArEnAV BA-TFD AV 75.91 44.31 77.64 58.29 72.21 45.21
Frame-level AV-1M + ArEnAV BA-TFD+ AV 79.97 27.44 84.20 36.47 72.89 34.56

Temporal localization:

Results on ArEnAv, AV-1M and LAV-DF. The low performance on ArEnAV demonstrates the data complexity in CSW settings.

Method Dataset [email protected] [email protected] [email protected] AR@50 AR@20 AR@10
BA-TFD LAV-DF 79.15 38.57 0.24 64.18 60.89 58.51
AV-1M 37.37 6.34 0.02 45.55 35.95 30.66
ArEnAV 2.42 0.55 0.01 22.30 3.41 2.54
BA-TFD+ LAV-DF 96.30 84.96 4.44 80.48 79.40 78.75
AV-1M 44.42 13.64 0.03 48.86 40.37 34.67
ArEnAV 3.74 1.10 0.04 30.75 4.55 3.05

Human Evaluation:

User study results show that the deepfake detection and localization in multilingual CSW videos is non-trivial for human observers.

Metric Value
Accuracy 60 %
AP @ 0.1 8.35
AP @ 0.5 0.79
AR @ 1 1.38

Citation 

@misc{kuckreja2025tellhabibirealfake,
      title={Tell me Habibi, is it Real or Fake?}, 
      author={Kartik Kuckreja and Parul Gupta and Injy Hamed and Thamar Solorio and Muhammad Haris Khan and Abhinav Dhall},
      year={2025},
      eprint={2505.22581},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.22581}, 
}
Downloads last month
72