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

How face recognition and deepfake detection are measured

A score is one field on the record. It counts only when the test set, the protocol and the date beside it are filled in and match the source.

Face Benchmark editorial mark

How a number is read

No. 000 001
AModelXceptionThe network whose output was scored.
BPaper and venueDeepfakeBench, NeurIPS 2023Where the number was printed, Table 3.
CTrain setFaceForensics++ c23What the model learned from.
DTest setCeleb-DF v2What it was measured on. The field to check first.
EProtocolCross-datasetUnseen fakes, c23 training compression.
FMetricFrame-level AUCWhich ruler, scored per frame.
GValue0.7365Void without fields C to F.CHECKED
HDate checked8 Nov 2026When we last matched it to the paper.
Machine-readable stripThe last character checks the two lines. Copy one field wrong and it stops matching.

Machine-readable strip: result record, Xception, DeepfakeBench, NeurIPS 2023, FaceForensics++ c23, Celeb-DF v2, Cross-dataset, Frame-level AUC, 0.7365, 8 Nov 2026, check digit 9.

Every field on this record is copied from Table 3 of the DeepfakeBench paper (NeurIPS 2023), read 8 November 2026. The same row appears on the results page.

02

What this site covers

Face Benchmark explains how face recognition and deepfake detection are measured. It describes the public datasets, how each was collected and under what licence; the metrics used to score models; and published results copied from papers and official leaderboards, each with its test set, protocol and the date it was checked. It hosts no datasets, no face images and no detection tool, and it does not say whether any particular video is real.

  • 01DatasetsCollection, consent, licence and the access route.
  • 02MetricsWhat AUC, EER and TAR at FAR measure, and what they hide.
  • 03Published resultsNumbers from the source, one click away, with conditions.
  • 04LimitsWhere a benchmark stops and a single clip begins.

This domain's history

WEBFACE260M SITE, 2021 TO 2025

From 2021 to 2025 this domain was the WebFace260M benchmark site. This site is not the WebFace260M project and is not affiliated with its authors, Tsinghua University or XForwardAI. Two pages keep the record for readers arriving from old papers and READMEs.

The ICCV 2021 masked face recognition challenge WebFace260M and WebFace42M
04

Deepfake detection benchmarks

Three registers: the test sets, the scores reported on them, and what a person holding one clip can check.

D.1Datasets

Where the test sets come from and how to get them.

  • Public deepfake datasets compared

    The public sets side by side: year, methods, rights, licence, access.

  • FaceForensics++

    Its manipulation methods, compression levels and access route.

  • Celeb-DF

    What it set out to fix and what its test split holds.

  • DFDC

    The challenge dataset: splits, licence and how it was scored.

  • DeepfakeBench

    One codebase for training and scoring detectors alike.

D.2Results and metrics

What a score means and when it holds.

D.3Checking media

What a person can check, in order of evidence.

See all deepfake detection pages
  1. FR 01Face recognition benchmarksThe test sets and protocols face recognition is scored on.
  2. FR 02Face verification metricsTAR at a fixed FAR, ROC curves and what 1:1 means.
  3. FR 03NIST's evaluationsHow the FRTE series tests algorithms and reports results.
  4. FR 04Demographic differentialsHow error rates are compared across groups, and the limits of that.
  5. FR 05Datasets and consentHow large face sets were collected, and the cases that followed.
06

How we report numbers

Source

The paper or official leaderboard the number came from, linked one click away, with the table it sits in.

Test set

The named test set and split. A number without one is not printed.

Protocol

In-dataset or cross-dataset, frame or video level, and the compression the clips were scored at.

Date

When the paper appeared and when we last matched our copy against it.

Numbers are copied, not rerun. If a field is missing from the source, the cell says so rather than guessing. Corrections are welcome and are dated where they change a figure. How Face Benchmark reports numbers.

No benchmark can tell you whether one particular video is real. It tells you how a method did on a test set, under a protocol, on a date.

07

Latest checks

  1. Published detector results: DeepfakeBench rows matched against Table 3 of the paper.

    All within-domain and cross-domain frame-level AUC values confirmed against arXiv 2307.01426; the repository README carries the same figures.

  2. Public deepfake datasets compared: dataset counts read from each paper and official repository.

    FaceForensics++, Celeb-DF, DFDC, DeeperForensics-1.0, WildDeepfake, ForgeryNet, DF40 and Deepfake-Eval-2024 rows dated.

  3. WebFace260M and WebFace42M: WebFace260M figures read from the CVPR 2021 paper.

    Identity and image counts, the CAST pipeline and the FRUITS protocol confirmed against arXiv 2103.04098.