Consent in deepfake datasets: which sets asked the people shown

This page belongs to the deepfake detection hub and fills in the rights column of the public deepfake datasets table one set at a time. The face recognition side of the same question, including the withdrawn training sets and the Clearview AI rulings, is covered on the face datasets and consent page and is not repeated here.

01Why consent is a benchmark question

A deepfake dataset is made of two things: real footage of people, and manipulated copies of that footage. Both show a real person's face, and the manipulated copy shows it doing something the person never did. Whether that person agreed changes who may hold the data, whether it can be redistributed, and whether it will still exist in five years, which is why the status column on the datasets table matters as much as the counts. DeeperForensics-1.0's authors stated the problem plainly when explaining why they filmed their own actors: "the lack of official informed consents from the video subjects precludes the use of these videos, even for non-commercial purposes" (arXiv 2001.03024).

The field's two most cited sets took the other route. FaceForensics++ (2019) and Celeb-DF (2020) were built from videos downloaded from YouTube, and their documentation says so. Neither claims consent; one offers removal on request. The sections below take each set's own paper or repository as the record and quote what it says, in the order consented, downloaded, found.

02The table: source, consent, licence, takedown

Every cell is copied from the dataset's paper or repository README, read on 9 November 2026. "None stated" means the document read does not address the point.

03Which deepfake datasets recorded consent?

The Deepfake Detection Challenge dataset is the largest consented set by a wide margin. Its paper states that it holds "over 100,000 total clips sourced from 3,426 paid actors" and that "all recorded subjects agreed to participate in and have their likenesses modified during the construction of the face-swapped dataset" (arXiv 2006.07397). The real and the fake halves of the set therefore share the same consent. The DFDC dataset page covers its splits and scoring.

DeeperForensics-1.0 took the same route at smaller scale and documented it in more detail. Its paper reports that "100 paid actors give consents to the use and manipulation of their faces by signing a formal agreement", that the actors were 55 men and 45 women from 26 countries aged 20 to 45, and that their faces were swapped onto 1,000 driving videos to make the fakes (arXiv 2001.03024). The driving videos are the set's own recordings, so no found footage enters the fakes either.

The Deep Fake Detection dataset contributed by Google and Jigsaw in September 2019 is the third. Google's announcement says the company "worked with paid and consenting actors to record hundreds of videos" and swapped faces between randomly paired actors (Google AI blog, 2019); the FaceForensics++ repository, which hosts it, counts over 3,000 manipulated videos from 28 actors.

04Which datasets used YouTube without consent?

FaceForensics++ is the field's default training set and its 1,000 source sequences came from 977 YouTube videos. The paper's appendix describes the collection: "we chose to collect videos in the wild, more specifically from YouTube", using the YouTube-8m tags "face", "newscaster" and "newsprogram" and search terms such as "interview" and "blog", keeping videos of 480p or better (arXiv 1901.08971). Nothing in the paper, the README or the terms of use records consent from the people in those videos. The set is distributed by Google form under the FaceForensics Terms of Use, which restrict it to "non-commercial research and educational purposes", make the researcher responsible for "any copies of copyrighted Images, Videos" they create, and let the university "terminate Researcher's access to the Database at any time". The terms contain no removal clause.

Celeb-DF took 590 videos of 59 public figures from YouTube, plus 300 further YouTube videos used as additional real footage, and produced 5,639 face swaps from them. Its README carries a privacy statement that is unusually direct: the original videos "are obtained from the Internet which are not property of the authors or the authors' affiliated institutions", and "If you feel uncomfortable about your identity shown in this dataset, please contact us and we will remove corresponding information from the dataset" (Celeb-DF repository). That is the only takedown route stated by any set in the table.

Both sets were built this way for a technical reason. Manipulation methods of 2018 and 2019 needed frontal, unoccluded, well lit faces, and public interview and news footage supplied them in quantity. The consent position is absent from both records, and this page reports that absence rather than inferring an answer.

05What about datasets collected from the internet?

WildDeepfake is a different case again. Its 707 deepfake videos were "collected completely from the internet", and its real videos were collected the same way (arXiv 2101.01456). The people in the fakes are whoever the anonymous makers chose to swap, and the people in the real clips are whoever posted them. No consent is possible after the fact, and the authors do not claim any. The same holds for the 2024 in-the-wild collections described on the datasets table: their value is that they contain what actually circulated, and the price of that value is that nobody in them agreed.

A found-footage set documents a distribution; it cannot document permission, and its card can only say where the clips came from and under what terms they are shared.

06How does this change what a detector score means?

Consent does not change a frame, so a detector's AUC on DFDC is as valid a number as its AUC on FaceForensics++. What changes is the set's future. MegaFace and MS-Celeb-1M, two of the large sets listed among the list of face recognition benchmarks, were withdrawn after questions about the people in them, and a score on a withdrawn set cannot be reproduced by anyone who did not already hold a copy. The faces are the same kind of data whether the task is recognising who someone is or deciding whether a frame was manipulated, and the same law applies to both. Of the deepfake sets, the three consented ones are the ones whose documentation would survive the same scrutiny; the two YouTube sets rely on research terms of use and, in Celeb-DF's case, on removal by request.

The second consequence is scope. DFDC's paper records that a model trained only on its consented actors "can generalize to real in-the-wild Deepfake videos", a 2020 result on 2020 fakes. A consented set can be large, varied and useful; what it cannot be, by construction, is a sample of the fakes that circulate, because those are made of people who did not sign anything.