Face BenchmarkHow face recognition and deepfake detection are measured, from the papers.
Menu

Deepfake detector results from published papers: in-dataset and cross-dataset scores, with dates

The Face Benchmark editorsPublishedLast checked

Short answer

119 words

The highest cross-dataset score in the shared DeepfakeBench evaluation is SPSL's frame-level AUC of 0.7650 on Celeb-DF v2 after training on FaceForensics++ c23 (NeurIPS 2023, Table 3), against 0.9610 when the same detector is tested on the FaceForensics++ split it was trained on. Ten of the fifteen detectors in that table score above 0.93 at home, and all fifteen score between 0.55 and 0.95 on unseen sets, with cross-dataset averages from 0.6551 (Meso4) to 0.7875 (SPSL). On fakes that circulated online in 2024, the best off-the-shelf open-source video model scored 0.63 AUC and the best commercial one 0.79 (Deepfake-Eval-2024). The table below gives each value with its train set, test set, protocol, metric and the date it was checked.

TBLPublished results

Every value is copied from the paper named in its row and matched against it on the date in the last column. Rows are grouped by protocol: in-dataset first, then cross-dataset, then benchmarks built from fakes that circulated. The strip under each row carries the same fields in machine-readable form, ending in a check digit.

Deepfake detector results from published papersFrame-level AUC unless the metric column says otherwise. For log loss, lower is better.
01Model02Paper and venue03Year04Train set05Test set06Protocol and compression07Metric08Value09Verdict10Best for11Main limitation12Code available13Date checked
GRP INDIn-datasetTrained and tested on splits of the same dataset.
XceptionNetFaceForensics++, ICCV 2019 (Table 1)2019FF++ (all four methods)FF++ test splitIn-dataset, raw (uncompressed)Binary accuracy99.26%The ceiling: near-perfect on uncompressed clips of the same methods.Showing what compression removesRaw video is rare outside a lab.Yes2026-11-08
Strip: XceptionNet, FaceForensics++, ICCV 2019 (Table 1), 2019, FF++ (all four methods), FF++ test split, In-dataset, raw (uncompressed), Binary accuracy, 99.26%, The ceiling: near-perfect on uncompressed clips of the same methods., Showing what compression removes, Raw video is rare outside a lab., Yes, 2026-11-08, check digit 1.
XceptionNetFaceForensics++, ICCV 2019 (Table 1)2019FF++ (all four methods)FF++ test splitIn-dataset, LQ (c40)Binary accuracy81.00%Heavy compression alone costs 18 points on the same fakes.Reading the compression columnAccuracy depends on the real to fake balance of the split.Yes2026-11-08
Strip: XceptionNet, FaceForensics++, ICCV 2019 (Table 1), 2019, FF++ (all four methods), FF++ test split, In-dataset, LQ (c40), Binary accuracy, 81.00%, Heavy compression alone costs 18 points on the same fakes., Reading the compression column, Accuracy depends on the real to fake balance of the split., Yes, 2026-11-08, check digit 5.
XceptionDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23FF++ c23 testIn-dataset, c23Frame-level AUC0.9637The reference baseline most later detectors are measured against.A baseline for any FF++ comparisonFrame-level, so one clip contributes many scores.Yes2026-11-08
Strip: Xception, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, FF++ c23 test, In-dataset, c23, Frame-level AUC, 0.9637, The reference baseline most later detectors are measured against., A baseline for any FF++ comparison, Frame-level, so one clip contributes many scores., Yes, 2026-11-08, check digit 5.
UCFDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23FF++ c23 testIn-dataset, c23Frame-level AUC0.9705The highest in-dataset score in the table; the within-domain average 0.9527 is also the best.Seen fakes at c23A 0.007 lead over Xception in-dataset does not carry to unseen sets.Yes2026-11-08
Strip: UCF, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, FF++ c23 test, In-dataset, c23, Frame-level AUC, 0.9705, The highest in-dataset score in the table; the within-domain average 0.9527 is also the best., Seen fakes at c23, A 0.007 lead over Xception in-dataset does not carry to unseen sets., Yes, 2026-11-08, check digit 4.
SPSLDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23FF++ c23 testIn-dataset, c23Frame-level AUC0.9610Slightly below Xception in-dataset; the cross-dataset rows are where it stands out.Pairing with its cross-dataset rowsPhase-spectrum cues weaken under recompression.Yes2026-11-08
Strip: SPSL, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, FF++ c23 test, In-dataset, c23, Frame-level AUC, 0.9610, Slightly below Xception in-dataset; the cross-dataset rows are where it stands out., Pairing with its cross-dataset rows, Phase-spectrum cues weaken under recompression., Yes, 2026-11-08, check digit 4.
EfficientNet-B4DeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23FF++ c23 testIn-dataset, c23Frame-level AUC0.9567A plain classifier within 0.01 of the purpose-built methods on seen fakes.Judging whether a new method adds anythingNo deepfake-specific design to analyse.Yes2026-11-08
Strip: EfficientNet-B4, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, FF++ c23 test, In-dataset, c23, Frame-level AUC, 0.9567, A plain classifier within 0.01 of the purpose-built methods on seen fakes., Judging whether a new method adds anything, No deepfake-specific design to analyse., Yes, 2026-11-08, check digit 0.
XceptionDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23FF++ c40 testIn-dataset methods, heavier compression (c40)Frame-level AUC0.8261Same methods, same faces, more compression: a 0.14 drop.Separating compression from generalisationc40 is still a lab setting, not platform re-encoding.Yes2026-11-08
Strip: Xception, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, FF++ c40 test, In-dataset methods, heavier compression (c40), Frame-level AUC, 0.8261, Same methods, same faces, more compression: a 0.14 drop., Separating compression from generalisation, c40 is still a lab setting, not platform re-encoding., Yes, 2026-11-08, check digit 7.
Winning entry (S. Seferbekov)DFDC dataset paper, 2020 (Table 2)2020DFDC training setDFDC private test set (10,000 clips)Challenge, mixed perturbationsLog loss (lower is better)0.4279The winning competition entry on a private set half made of organic internet clips.A realistic ceiling for 2020 methodsLog loss is not comparable with AUC; the DFDC-only part scored 0.1983, the real-only part 0.6605.Yes2026-11-08
Strip: Winning entry (S. Seferbekov), DFDC dataset paper, 2020 (Table 2), 2020, DFDC training set, DFDC private test set (10,000 clips), Challenge, mixed perturbations, Log loss (lower is better), 0.4279, The winning competition entry on a private set half made of organic internet clips., A realistic ceiling for 2020 methods, Log loss is not comparable with AUC; the DFDC-only part scored 0.1983, the real-only part 0.6605., Yes, 2026-11-08, check digit 0.
GRP XDSCross-datasetTrained on one dataset, tested on another it never saw.
XceptionDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23Celeb-DF v2Cross-dataset, c23Frame-level AUC0.7365The field's most quoted gap: 0.9637 at home, 0.7365 on Celeb-DF v2.The reference cross-dataset numberCeleb-DF v2 is one swap method; the drop differs on other sets.Yes2026-11-08
Strip: Xception, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, Celeb-DF v2, Cross-dataset, c23, Frame-level AUC, 0.7365, The field's most quoted gap: 0.9637 at home, 0.7365 on Celeb-DF v2., The reference cross-dataset number, Celeb-DF v2 is one swap method; the drop differs on other sets., Yes, 2026-11-08, check digit 7.
XceptionDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23DFDCCross-dataset, c23Frame-level AUC0.7077Lower again on the DFDC set, which adds perturbations and more methods.Reading how test-set difficulty moves the numberDFDC frames are heavily perturbed by design.Yes2026-11-08
Strip: Xception, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, DFDC, Cross-dataset, c23, Frame-level AUC, 0.7077, Lower again on the DFDC set, which adds perturbations and more methods., Reading how test-set difficulty moves the number, DFDC frames are heavily perturbed by design., Yes, 2026-11-08, check digit 5.
XceptionDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23FaceShifterCross-dataset, c23Frame-level AUC0.6249A newer swap method on the same source videos nearly halves the margin over chance.Showing that the generator matters more than the facesFaceShifter is one method; results vary across the eight sets.Yes2026-11-08
Strip: Xception, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, FaceShifter, Cross-dataset, c23, Frame-level AUC, 0.6249, A newer swap method on the same source videos nearly halves the margin over chance., Showing that the generator matters more than the faces, FaceShifter is one method; results vary across the eight sets., Yes, 2026-11-08, check digit 2.
UCFDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23Celeb-DF v2Cross-dataset, c23Frame-level AUC0.7527Better than Xception on Celeb-DF v2 by 0.016; its cross-dataset average over eight sets is 0.7801.A 2023 method's unseen-set scoreStill 0.22 below its own in-dataset score.Yes2026-11-08
Strip: UCF, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, Celeb-DF v2, Cross-dataset, c23, Frame-level AUC, 0.7527, Better than Xception on Celeb-DF v2 by 0.016; its cross-dataset average over eight sets is 0.7801., A 2023 method's unseen-set score, Still 0.22 below its own in-dataset score., Yes, 2026-11-08, check digit 2.
SPSLDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23Celeb-DF v2Cross-dataset, c23Frame-level AUC0.7650The best Celeb-DF v2 score in the table, and the best cross-dataset average, 0.7875.Unseen fakes, from the published tableBest among these detectors is still a quarter of the way to chance.Yes2026-11-08
Strip: SPSL, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, Celeb-DF v2, Cross-dataset, c23, Frame-level AUC, 0.7650, The best Celeb-DF v2 score in the table, and the best cross-dataset average, 0.7875., Unseen fakes, from the published table, Best among these detectors is still a quarter of the way to chance., Yes, 2026-11-08, check digit 8.
SPSLDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23Celeb-DF v1Cross-dataset, c23Frame-level AUC0.8150The earlier, smaller Celeb-DF is easier for every detector than v2.Reading version differences in a test setv1 has 795 fakes; v2 has 5,639.Yes2026-11-08
Strip: SPSL, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, Celeb-DF v1, Cross-dataset, c23, Frame-level AUC, 0.8150, The earlier, smaller Celeb-DF is easier for every detector than v2., Reading version differences in a test set, v1 has 795 fakes; v2 has 5,639., Yes, 2026-11-08, check digit 3.
EfficientNet-B4DeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23DFDCCross-dataset, c23Frame-level AUC0.6955The plain classifier drops the most on DFDC, 0.26 below its in-dataset score.A floor for cross-dataset claimsNot designed for generalisation.Yes2026-11-08
Strip: EfficientNet-B4, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, DFDC, Cross-dataset, c23, Frame-level AUC, 0.6955, The plain classifier drops the most on DFDC, 0.26 below its in-dataset score., A floor for cross-dataset claims, Not designed for generalisation., Yes, 2026-11-08, check digit 4.
FWADeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23DeeperForensics-1.0Cross-dataset, c23Frame-level AUC0.9334The best DeeperForensics-1.0 score in the table, from a 2019 warping-artefact method.Face-warping artefactsIts average across the eight unseen sets is 0.7239, below Xception.Yes2026-11-08
Strip: FWA, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, DeeperForensics-1.0, Cross-dataset, c23, Frame-level AUC, 0.9334, The best DeeperForensics-1.0 score in the table, from a 2019 warping-artefact method., Face-warping artefacts, Its average across the eight unseen sets is 0.7239, below Xception., Yes, 2026-11-08, check digit 2.
Face X-rayDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23DeeperForensics-1.0Cross-dataset, c23Frame-level AUC0.5531Near chance on a set whose swaps leave no blending boundary of the kind it learns.Showing that a cue can be dataset-specificReproduced by the benchmark authors; the original code was unpublished.Reproduced2026-11-08
Strip: Face X-ray, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, DeeperForensics-1.0, Cross-dataset, c23, Frame-level AUC, 0.5531, Near chance on a set whose swaps leave no blending boundary of the kind it learns., Showing that a cue can be dataset-specific, Reproduced by the benchmark authors; the original code was unpublished., Reproduced, 2026-11-08, check digit 2.
Face X-rayDeepfakeBench, NeurIPS 2023 (Table 3)2023FF++ c23FaceShifterCross-dataset, c23Frame-level AUC0.6553The best FaceShifter score in the table, and still the lowest column for most detectors.Blending-boundary swapsCross-dataset average 0.6985, the lowest among the serious methods.Reproduced2026-11-08
Strip: Face X-ray, DeepfakeBench, NeurIPS 2023 (Table 3), 2023, FF++ c23, FaceShifter, Cross-dataset, c23, Frame-level AUC, 0.6553, The best FaceShifter score in the table, and still the lowest column for most detectors., Blending-boundary swaps, Cross-dataset average 0.6985, the lowest among the serious methods., Reproduced, 2026-11-08, check digit 0.
GRP WLDIn the wildTested on fakes that circulated online, where the generator and the compression are unknown.
GenConViTDeepfake-Eval-2024, 2025 (Table 3)2025Its original training dataDeepfake-Eval-2024 videoOff the shelf, platform compressionAUC0.63The best off-the-shelf open-source video model on circulated 2024 fakes, down from 0.96 on its own data.The size of the lab-to-wild gapFine-tuning on part of the set raises it to 0.82 (Table 4), which is a different protocol.Yes2026-11-08
Strip: GenConViT, Deepfake-Eval-2024, 2025 (Table 3), 2025, Its original training data, Deepfake-Eval-2024 video, Off the shelf, platform compression, AUC, 0.63, The best off-the-shelf open-source video model on circulated 2024 fakes, down from 0.96 on its own data., The size of the lab-to-wild gap, Fine-tuning on part of the set raises it to 0.82 (Table 4), which is a different protocol., Yes, 2026-11-08, check digit 3.
Best commercial video model (unnamed in the paper)Deepfake-Eval-2024, 2025 (Table 5)2025UndisclosedDeepfake-Eval-2024 videoOff the shelf, platform compressionAUC0.79Commercial video models ranged from 0.64 to 0.79 on the same set.A ceiling for 2025 off-the-shelf tools on real fakesThe paper does not name the vendors; nothing here can be matched to a product.No2026-11-08
Strip: Best commercial video model (unnamed in the paper), Deepfake-Eval-2024, 2025 (Table 5), 2025, Undisclosed, Deepfake-Eval-2024 video, Off the shelf, platform compression, AUC, 0.79, Commercial video models ranged from 0.64 to 0.79 on the same set., A ceiling for 2025 off-the-shelf tools on real fakes, The paper does not name the vendors; nothing here can be matched to a product., No, 2026-11-08, check digit 9.
DFDC winning modelDFDC dataset paper, 20202020DFDC training setReal in-the-wild deepfake videos collected by the authorsCross-source, organic clipsROC-AUC (real videos only)0.734A 2020 model trained only on paid-actor fakes kept 0.734 AUC on organic deepfakes.Evidence that DFDC training generalised in 2020The in-the-wild set is small and from 2020 generators.Yes2026-11-08
Strip: DFDC winning model, DFDC dataset paper, 2020, 2020, DFDC training set, Real in-the-wild deepfake videos collected by the authors, Cross-source, organic clips, ROC-AUC (real videos only), 0.734, A 2020 model trained only on paid-actor fakes kept 0.734 AUC on organic deepfakes., Evidence that DFDC training generalised in 2020, The in-the-wild set is small and from 2020 generators., Yes, 2026-11-08, check digit 0.
  • FBRresult record
  • Amodel
  • Bpaper
  • Cyear
  • Dtrain set
  • Etest set
  • Fprotocol
  • Gmetric
  • Hvalue
  • Idate checked
  • #check digit over the line

01How to read the table

This is the results page of the deepfake detection benchmark hub. The datasets the rows refer to are described on public deepfake datasets compared, and the metrics on how detectors are scored.

Frame-level AUC of four detectors trained on FaceForensics++ c23, tested on the FF++ c23 split and on Celeb-DF v2, from DeepfakeBench Table 3Grouped bars. Xception 0.9637 in-dataset and 0.7365 on Celeb-DF v2; UCF 0.9705 and 0.7527; SPSL 0.9610 and 0.7650; EfficientNet-B4 0.9567 and 0.7487. Source: DeepfakeBench, NeurIPS 2023, Table 3, read 8 Nov 2026.0.00000.25000.50000.75001.0000FF++ c23 test (in-dataset)Celeb-DF v2 (cross-dataset)Xception0.96370.7365UCF0.97050.7527SPSL0.96100.7650EfficientNet-B40.95670.7487
Figure 1The same four detectors, same training set, two test sets. The left bar of each pair is the in-dataset score, the right bar the cross-dataset score on Celeb-DF v2. Source: DeepfakeBench Table 3 (arXiv 2307.01426), frame-level AUC, read 8 Nov 2026.

Each row is one number from one paper. Read across from the model to the value, and treat the columns between them as part of the number: a value without its test set, protocol and metric is not a result. The verdict column is our short reading of what the row shows, not a ranking, and the "best for" column says what question the row answers.

In-dataset rows

The detector was trained and tested on splits of the same dataset. These rows show the ceiling: how well a method does on fakes like the ones it learned from. They are usually the highest numbers a paper reports, and the first row, 99.26% binary accuracy on uncompressed FaceForensics++, is the clearest case.

Cross-dataset rows

The detector was trained on one dataset and tested on another it never saw. These rows are closer to how a detector meets new material, and the drop from the in-dataset row is often the most useful figure on the page. The cross-dataset generalisation page works through the drops model by model.

AUC is not accuracy

Most rows report the area under the ROC curve at frame level, where 0.5 is chance and 1.0 is perfect separation across every threshold. Two rows report binary accuracy, which depends on one chosen threshold and on how many fakes the test split holds, and one reports log loss, where lower is better. None of the three can be compared with another directly. The metrics page defines each, and why a frame-level AUC and a video-level AUC on the same set are different measurements.

Why most rows come from one benchmark

Fourteen of the rows are taken from a single table, Table 3 of the DeepfakeBench paper, because it is the only published source where many detectors were run through the same preprocessing, the same splits and the same metric. Numbers from separate papers, each with its own face crops and frame sampling, cannot be put in one column honestly; DeepfakeBench exists because they were. The in-the-wild rows come from the two papers that built their test sets from circulated fakes, and the challenge row from the DFDC dataset paper.

02Method and sources

Values are copied from the paper or an official leaderboard, never rerun or estimated. When a paper reports several settings, we take the one its authors present as the main result and say so in the protocol column. Where a repository README and the paper both carry a table, we read the paper; for DeepfakeBench the README states that its results table is cited from the paper, and the two agree.

Which sources qualify
Peer-reviewed papers, preprints with released code, and leaderboards or test sets run by a dataset's maintainers. Claims without a stated test set are not included, and commercial products are never named or scored here; the one commercial row is the anonymous best model in a published study.
Conflicts
Where two sources report different values for the same setting, both rows are shown with their sources and the verdict column notes the disagreement. No conflicts stand in the current table.
Last check
Every row was matched against its source on the date in its own column. The table as a whole was last checked on 8 November 2026, and the DeepfakeBench repository said on that date that an expanded evaluation "will update the below table soon"; when the paper's table changes, the rows will be re-checked and re-dated.

Reading a paper that is not in this table is covered on how to read a detection paper; testing a detector of your own so that the result can sit beside these rows is the evaluate a detector protocol.

A value holds only for the train set, test set, protocol, metric and date in its own row. Newer generators lower every score here, and no score says whether one particular video is real.

FAQQuestions

Which deepfake detector is the most accurate?

There is no single answer, because the rows were measured under different conditions. Within the one shared evaluation in the table, UCF has the highest in-dataset AUC (0.9705) and SPSL the highest cross-dataset average (0.7875), both trained on FaceForensics++ c23 and scored at frame level. Neither number says how either detector would do on fakes made in 2026.

Why are the cross-dataset scores lower?

A detector learns the traces left by the manipulation methods in its training data. On a dataset made with other methods, at other compression levels, many of those traces are missing, so scores fall. The size of the drop is often more informative than either number alone.

Can I compare two values in the table directly?

Only when their train set, test set, protocol, compression and metric all match. The DeepfakeBench rows can be compared with each other because the benchmark fixed all five. A frame-level AUC cannot be compared with a binary accuracy or a log loss.

Where do the numbers come from?

Each value is copied from the paper or the official leaderboard named in the Sources list and in its own row. We do not rerun models or estimate missing values; a field the source does not state is marked as not stated.

What does the date checked mean?

It is the day we last matched the row against its source. Papers are revised and repository tables change, so a figure should be cited with its date.

Why is there no overall ranking?

Rows were measured under different conditions, so a single ordering would compare unlike things. The verdict column summarises what each row shows on its own terms.

Why are commercial detectors not listed?

The table reports published research results. The one commercial row is the anonymous best model in a peer-reviewed study that did not name vendors, included because it is the only published measurement of commercial tools on circulated fakes. This site does not rank products.

What does the in-the-wild group measure?

Scores on fakes collected after they circulated, where the generator and the platform compression are unknown. Deepfake-Eval-2024 reports that open-source models' AUC fell by about half compared with their original benchmarks. The in the wild page covers those sets.

How do I report a wrong value?

Send the row, the value you believe is correct and the page or table in the source where it appears. Corrections that change a figure are dated on the page. The address is on the contact page.

SRCSources

  1. S1
    DeepfakeBench: A Comprehensive Benchmark of Deepfake Detection, Table 3(external)Yan, Zhang, Yuan, Lyu, Wu. NeurIPS 2023 Datasets and Benchmarks
    Checked
  2. S2
    FaceForensics++: Learning to Detect Manipulated Facial Images, Table 1(external)Rössler, Cozzolino, Verdoliva, Riess, Thies, Nießner. ICCV 2019
    Checked
  3. S3
    The DeepFake Detection Challenge (DFDC) Dataset, Table 2(external)Dolhansky, Bitton, Pflaum, Lu, Howes, Wang, Ferrer. 2020
    Checked
  4. S4Checked
  5. S5Checked