Deepfake detection datasets by year: a 2018 to 2025 timeline

This page orders the sets in time for the deepfake detection hub. The side by side dataset table compares the same sets by rights, methods and compression; this page asks what changed from one year to the next.

01Why order deepfake detection datasets by year?

Because each set was built to fix what the previous one lacked, and the sequence shows what the field thought the problem was at each point. The Celeb-DF authors drew the first line: they "categorize UADFV, DF-TIMIT, and FF-DF as the first generation of DeepFake datasets, while DFD, DFDC, and the proposed Celeb-DF datasets are the second generation", the second generation having more videos and better visual quality (arXiv 1909.12962). A third wave in 2020 and 2021 added scale, consent and new modalities, and a fourth from 2024 stopped making fakes in the lab and started cataloguing generators or collecting what circulated.

Counts in the table are as each paper or comparison table states them; where two sources give different figures for one set, the difference is listed in the section on differing numbers. No file, frame or face from any set is hosted here.

02The timeline: 14 datasets, 2018 to 2025

Release dates for the first six rows are from Celeb-DF Table 1; later rows take the month of first arXiv submission. DF-TIMIT's subject count is 43 in the ForgeryNet table and 32 in the WildDeepfake and FakeAVCeleb tables, so it is not stated. DF40 reports technique counts (10 face swapping, 13 face reenactment, 12 entire face synthesis, 5 face editing) rather than one video total (arXiv 2406.13495).

03What changed between 2018 and 2019?

Scale by a factor of forty, and the first standard methods. UADFV was 98 videos made with one early consumer application and is cited to the head-pose detection paper of Yang, Li and Lyu (arXiv 1811.00661); DF-TIMIT was 640 fakes from VidTIMIT at two qualities, built to show that face recognition systems accepted the swaps (arXiv 1812.08685). FaceForensics++ in January 2019 fixed the pattern every later set followed: 1,000 source videos, four named manipulation methods applied to each, three compression levels, and fixed splits (arXiv 1901.08971).

The second generation, in the Celeb-DF authors' term, answered quality. Celeb-DF's stated motivation was that existing fakes "suffer from low visual quality and do not resemble DeepFake videos circulated on the Internet". DFD in September 2019 was the first set made with consenting actors, over 3,000 manipulated videos from 28 of them, and the DFDC Preview a month later carried an agreement from every recorded person to the manipulation of their likeness (arXiv 1910.08854).

04What did 2020 and 2021 add?

Three things: order-of-magnitude scale, consent as the default, and sources other than YouTube. DFDC's full release, over 100,000 clips from 3,426 paid actors with eight methods and nineteen perturbations, was the first set large enough to run a public competition on (arXiv 2006.07397). DeeperForensics-1.0 put 50,000 real videos of 100 paid actors against 10,000 fakes with extensive perturbations (arXiv 2001.03024). ForgeryNet reached 221,247 clips and 15 manipulation approaches with over 5,400 subjects (arXiv 2103.05630).

The other two 2021 sets changed who and what was in the frame. KoDF holds 175,776 fake and 62,166 real clips of 403 paid Korean participants, made with six synthesis models including two driven by audio, because its authors found the distribution of existing sets did not match Korean subjects (arXiv 2103.10094). FakeAVCeleb paired 500 real VoxCeleb2 videos with 19,500 fakes whose audio was also synthesised and lip-synced, the first set with fine-grained labels for both the face and the voice, by its own comparison table (arXiv 2108.05080). WildDeepfake, published at ACM Multimedia 2020, went the other way entirely: 707 videos collected from the internet with unknown methods, cut into 7,314 face sequences (arXiv 2101.01456).

05How did 2024 and 2025 change the question?

The fourth wave gave up on a single curated distribution. DF40's authors described FaceForensics++ as containing "out-of-date forgery techniques from the past four years" and instead generated fakes with 40 distinct techniques, running over 2,000 evaluations across four protocols and eight detectors; their second finding was that well-designed detectors "may not have obvious advantages over the baseline". Deepfake-Eval-2024 collected 45 hours of video, 56.5 hours of audio and 1,975 images that users had flagged on social media and a detection platform during 2024, from 88 websites in 52 languages, and found open-source detectors' AUC down by 50 percent for video against their original benchmarks (arXiv 2503.02857).

The two sets measure different things from the earlier ones. A 2019 set asks whether a detector can find its own generation of fakes; DF40 asks whether it survives forty generators; Deepfake-Eval-2024 asks whether it works on whatever the internet produced that year. The full DeepfakeBench cross-dataset grid shows how detectors trained on a 2019 set score on the 2018 to 2020 sets in this table, and the what a benchmark is page explains why a saturated set such as FaceForensics++ stops separating detectors.

06Why do the counts differ between sources?

Four cases in this table have two published figures. DFDC Preview is 4,119 fake videos in the DeepfakeBench README and 4,113 in the Celeb-DF and WildDeepfake comparison tables; this page uses 4,119 and names the other. DFD is 3,068 fake videos in the Celeb-DF and ForgeryNet tables and "over 3000" in the FaceForensics README; the tables give the exact count. DF-TIMIT's subjects are 43 or 32 depending on the comparison table, so no figure is given. Deepfake-Eval-2024's video total is 45 hours in the paper's abstract, 45.1 hours in its Table 1 and 44 hours in its repository README; this page uses the abstract figure.

Release dates also depend on the event chosen. Celeb-DF Table 1 gives month of release, which for FaceForensics++ is 2019.01 (the first arXiv version) although the paper appeared at ICCV in August 2019; WildDeepfake was published at ACM Multimedia 2020 but reached arXiv in January 2021. The table states which convention each row uses. Every face dataset in this list was built from footage of real people, and the consent position moved from none (first generation) to paid, agreeing participants (DFD onward) and back to found footage with no consent record (WildDeepfake, Deepfake-Eval-2024); the consent and dataset withdrawals page and the TSA checkpoint program page, where the subject consents at the podium, show the same question in face recognition.