InDeepFake: A Novel Multimodal Multilingual Indian Deepfake Video Dataset

Name: InDeepFake: A Novel Multimodal Multilingual Indian Deepfake Video Dataset
Creator: Ruchira Naskar

Citation Author(s):: Arnab Kumar Das (Indian Institute of Engineering Science and Technology, Shibpur, India)

Aritra Bose (Indian Institute of Engineering Science and Technology, Shibpur, India)

Priya Manohar (Indian Institute of Engineering Science and Technology, Shibpur, India)

Anurag Dutta (Indian Institute of Technology, Kharagpur, India)

Ruchira Naskar (Indian Institute of Engineering Science and Technology, Shibpur, India)

Rajat Subhra Chakraborty (Indian Institute of Technology, Kharagpur, India)
Submitted by:: Ruchira Naskar
Last updated:: Mon, 12/22/2025 - 05:09
DOI:: 10.21227/395c-wt73
Data Format:: *.mp4
Research Article Link:: Paper Link

992 views

Categories:

Keywords:

Deepfake dataset

Deepfake

deepfake detection

Indian deepfake dataset

Artificial Intelligence (AI); Media Content Generation; Cultural Identity; Deepfake Technology; Algorithmic Influence; Audience Trust; Digital Ethics; Arab Media; AI Literacy; Information Authenticity

Generative AI

Indian face database

multimodal dataset

CITE

AI-Powered Dataset Intelligence is available for this dataset exclusively to institutional subscribers.

Learn More

Abstract

Specifically created to meet the needs and limitations of deepfake detection technologies, InDeepFake is an innovative multilingual and multimodal audio-video deepfake face dataset for the Indian population. We adopted seven multimodal manipulation techniques for deepfake creation in this dataset. It presents a wide range of dialects and accents that represent the linguistic/ ethnic diversity of India as a nation.

All researches that use the dataset or any part of it must cite the following paper:

Arnab Kumar das, Aritra Bose, Priya Manohar, Anurag Dutta, Ruchira Naskar, and Rajat Subhra Chakraborty. "InDeepFake: A novel multimodal multilingual indian deepfake video dataset." Pattern Recognition Letters, vol. 197, pp. 16-23, 2025.

Paper Link: https://doi.org/10.1016/j.patrec.2025.07.002

We have open-sourced the code to implement the baseline methods at: https://github.com/arnabdasphd/InDeepFake.

Instructions:

The complete dataset is organized into two top-level directories: Real and Fake. The Real directory is further organized into two sub-directories, viz. Proprietary and Public. The Public sub-directory contains videos from YouTube, distributed over two age-group directories: (20-40) and (40-60); while the Proprietary directory presents all videos in the range of 20 to 40 years. Within each leaf-level directory, the real videos are stored, titled to clearly specify the subject’s gender and sample chronology. For example, a video titled id024_F, gives us the 24-th female subject’s video, and a video titled id137_M presents the 137-th male subject’s video. Age and character (real, proprietary, public, etc.) of the video can be obtained from the directory organization. The deepfake samples, are stored in the directory Fake. It further consists of six sub-directories: FSGAN, FaceSwap, DeepFaceLab, SV2TTS+Wav2Lip, E2TTS+Wav2Lip, and F5TTS+Wav2Lip. The deepfakes are titled based on their constituent reals. For example, a deepfake sample generated by FSGAN, synthesized by superimposing sample id024_F to sample id137_F, is stored into the sub-directory FSGAN, titled id024_F id137_F.