Table 1. Overview of prevalent deepfake detection datasets, summarizing their key attributes.

Dataset name	Origin	Size	Resolution	Diversity	Variety of attributes	Realism	Availability
CelebA	Academic	200k images	Various	Various	40 attribute annotations per image	High	Public
DeeperForensics-1.0	Academic	60,000 videos	1,920 × 1,080 pixels	100 actors from 26 countries of various skin tones and ages	Eight naturally expressed feelings	High	Public
Deepfake-TIMIT	Academic	640 videos	Low Quality: 64 × 64, High Quality: 128 × 128	Various	Face Swap	Low/High	Public
Celeb-DF	Academic	5,639 videos	256 × 256 pixels	59 celebrities of various ethnicities and ages	Face Swap	High	Public
FaceForensics (FF)	Academic	1,004 videos	Various	Various	Face-to-Face Reenactment	High	Public
UADFV	University of Albany	49 videos	294 × 500 pixels	Various	Physiological cues	Medium	Public
FaceForensics++ (FF++)	Academic	1,000 videos	VGA, HD, FullHD	Male 60%, Female 40%	Face	High	Public
Deepfake detection challenge (DFDC)	Facebook	5,000 videos	Various	74% female, 26% male, and diverse ethnicities	Face swap	High	Public
HOHA-based	Academic	600 videos	360 × 240 pixels	Various	Human actions	High	Public
FakeAVCeleb	Real YouTube videos of celebrities	Audio and Video	Various	Covers four racial backgrounds	Deepfake videos and corresponding synthesized cloned audios	High, with perfect lip-syncing	Public
Attack Agnostic	Combines two audio DeepFakes and one anti-spoofing dataset	Audio clips	Audio dataset, no visual resolution	Varied types of spoofing attacks	Audio DeepFakes designed to challenge detection algorithms	Designed to improve detection generalization	Public
ADD 2022	Audio Deep Synthesis Detection challenge	Covers a range of real-life scenarios	Audio dataset, no visual resolution	Real-life and challenging scenarios	Tracks for low-quality fake audio detection, partially fake audio	Aims to push boundaries of detection research	Public
TweepFake	Twitter, based on advancements in language modeling like GPT-2	25,572 tweets	Various	23 bots imitating 17 human accounts	Tweets generated using various techniques like Markov Chains, LSTM	Real tweets posted on Twitter	Public
ADBT	Twitter, focused on detecting bot-generated messages	Extended from existing work	Various	Generated tweets designed to mimic human writing	Tweets generated to test detection models against language models	Generated content difficult to distinguish from human-written	Public