DFDC 2020: what 2,114 teams scored on 10,000 unseen deepfake videos

The record enters the studies index as the first video deepfake benchmark in the index and the only one in which the organiser published the whole leaderboard. It is a challenge report by the dataset's builders, not a peer-reviewed test of commercial products (arXiv 2006.07397, abstract, read 9 November 2026).

The entrants were Kaggle teams, not products. The paper names the five prize winners and their methods, and the Meta post prints the top ten of the final leaderboard (Meta AI results post, leaderboard table, read 9 November 2026). None of the nine image or video detectors with a card in this index (Hive, AI or Not, Illuminarty, Sightengine, Deepware, Sensity, Reality Defender, TruthScan, Resemble Detect) appears among them; the entry ranked eighth, ID R&D, is a biometrics vendor without a card.

Teams and methods from Section 6.2 of the paper; leaderboard log losses from the Meta post, read 9 November 2026 (arXiv 2006.07397; Meta AI).

The data came from 3,426 paid actors who agreed to have their faces swapped, recorded for about 15 minutes each; 960 of them appear in the released dataset of 128,154 clips, of which 104,500 are unique face swaps (arXiv 2006.07397, Section 1 and Table 1, read 9 November 2026). Swaps were made with eight methods: two Deepfake Autoencoder sizes (DF-128 and DF-256), a morphable-mask nearest-neighbour method, Neural Talking Heads, FSGAN, a StyleGAN projection, a sharpening refinement pass and audio swaps (Table 1, footnote d).

Counts as printed in the paper (arXiv 2006.07397, Sections 3.5 and 6.1, read 9 November 2026). Augmenters were applied to about 70% of test videos and distractors to about 30%, including dog and flower-crown face filters (Section 3.6).

The ranking metric was log loss on the private set, chosen because the paper's preferred metric, a weighted precision that scales false positives by the ratio of real to fake videos in organic traffic, "can be extremely small and somewhat noisy" (Section 4). Every submission had to score all 10,000 private videos on one V100 GPU within 90 hours; most finished in under 10 hours, about 3.6 seconds per video (Section 6.1).

The headline figures are printed in Table 2 of the paper and the leaderboard table and body of the Meta post (arXiv 2006.07397, Table 2; Meta AI, read 9 November 2026).

Values as printed; the Table 2 column headed P@0.3 is captioned as precision at recall 0.5 in the same table (arXiv 2006.07397, Table 2, read 9 November 2026). Meta's post prints the winner's log loss as 0.42798 where the paper rounds to 0.4279.

The gap between the two halves of the private set is the result that matters for this index. The five winners held log losses of 0.17 to 0.22 on DFDC-style footage and 0.66 to 0.70 on organic internet videos, and the paper says "on real videos, there was an expected performance drop" before citing the 0.753 average precision as evidence that DFDC training "allows a model to generalize to real videos" (Section 6.2). Meta's post states the same shift as a ranking change: the eventual winner had placed fourth on the public leaderboard, and the other four winners had placed 37th, 6th, 10th and 17th (Meta AI, read 9 November 2026).

Meta's post also records what the winners shared: all five used pretrained EfficientNet networks fine-tuned only on DFDC data, most chose the B7 variant, and "none of the top-performing solutions used digital forensics techniques, like sensor noise fingerprints" (Meta AI, read 9 November 2026). Facebook's own researchers entered but were ineligible for the $1 million prize pool and their final submission did not appear on the leaderboard "due to run time issues".

The paper states that the task is "extremely difficult and still an unsolved problem" (abstract), that a planned perceptual study of video quality was delayed by COVID-19, and that only 960 of roughly 3,500 recorded identities made it into the dataset for want of time and compute (Section 7). It ranks by log loss while arguing that precision under a one-in-a-million prevalence is the metric that matters (Section 4). Meta's post says the models were stress-tested afterwards by Oxford, the National University of Singapore and TU Munich, who "presented methods to trigger false positives and false negatives", without printing those results (Meta AI, read 9 November 2026).

Evident from the sources: the 65.18% and 82.56% figures appear only in Meta's post, not in the paper, which prints log loss and precision instead; the entrants are research teams, so no commercial detector is scored; and the videos are 2019 and 2020 face swaps from eight named methods, with no diffusion or modern video generators.

What this record does not say: a 65.18% (accuracy) score by the best of 2,114 teams on 10,000 half-organic videos in 2020 is a ceiling for that competition, not a measurement of any detector sold today.

No card takes a result line: the paper and the post name challenge teams, and none of the index's image or video cards (Hive, Sightengine, Deepware, Sensity, Reality Defender) is scored (Meta AI, leaderboard table, read 9 November 2026).

The record counts as a video-modality study with zero indexed detectors in the ranking of independent studies by modality, and the modalities feature lists it under video. It sets the published baseline that the claim vs evidence gap ranking measures vendor video claims against. The home page shows which AI detectors exist and who has tested them.

Related records: Ha and Passananti 2024 is the index's image test of a commercial detector; RAID image 2025 tests Hive and Sightengine against adversarial images; ASVspoof 5 is the equivalent anonymised challenge for voice.