Cross dataset deepfake detection results: the full DeepfakeBench grid

This page is the reference grid for the deepfake detection hub: every cross-dataset cell of the table in one place. The results page with its conditions carries the headline rows; the pages that rank the detectors and measure each one's drop take their numbers from here.

01Where the grid comes from

DeepfakeBench (NeurIPS 2023 Datasets and Benchmarks) trains every detector with the same preprocessing, the same face crops and the same frame sampling, then evaluates each on 14 test sets: six within the FaceForensics++ family and eight outside it. Table 3 of the paper prints the AUC for each pairing; the paper states that "all detectors are trained on FF-c23 and evaluated on other data" and that the benchmark "currently adopts the frame level evaluation" (arXiv 2307.01426). The eight cross-domain columns are the subject of this page. The framework itself is described on its own page.

Column abbreviations follow the paper. CDFv1 and CDFv2 are Celeb-DF versions 1 and 2; DF-1.0 is DeeperForensics-1.0; DFD is the Google and Jigsaw Deep Fake Detection set; DFDC and DFDCP are the Deepfake Detection Challenge full set and preview; Fsh is FaceShifter applied to the FaceForensics++ source videos; UADFV is the 2018 set of 49 real and 49 fake videos. Avg is the mean of the eight cells in the row.

02The full grid: 15 detectors by 8 unseen test sets

Frame-level AUC, all detectors trained on FaceForensics++ c23, from DeepfakeBench Table 3 (arXiv 2307.01426v2), read 9 November 2026. Type is the paper's grouping: Naive, Spatial or Frequency. The two Meso rows use a MesoNet backbone, CNN-Aug uses ResNet, EfficientB4 uses EfficientNet, X-ray uses HRNet, Capsule and Recce use their own networks, and every other row uses Xception.

03Which test set is hardest for every detector?

Read down the columns and one set stands out. FaceShifter is the lowest cell for 11 of the 15 detectors; its column runs from 0.5551 (FWA) to 0.6553 (X-ray), and no detector reaches 0.66 on it. DFDC is next: the column runs from 0.5560 (Meso4) to 0.7191 (UCF), with nine detectors between 0.69 and 0.72. The two hardest sets are hard for different reasons. FaceShifter is a single newer swap method applied to the same source videos the detectors trained on, so the real footage is familiar and only the forgery is new. DFDC is 3,426 paid actors recorded in new settings with eight methods and nineteen perturbations, so both the real and the fake footage are new.

At the other end, UADFV is the highest cell for 11 detectors and DeeperForensics-1.0 for the other 4 (Meso4, MesoIncep, Capsule, FWA). The UADFV column runs from 0.7150 to 0.9528 and DF-1.0 from 0.5531 (X-ray) to 0.9334 (FWA). UADFV is a 2018 set of 98 videos made with an early consumer application, and DeeperForensics-1.0 is one face swapping method applied to FaceForensics++ source footage; both are closer to the 2019 training distribution than DFDC or FaceShifter.

04Which detector reads best, column by column?

Each column has a leader, and the leaders are not the same detector. SPSL leads both Celeb-DF columns (0.8150 and 0.7650). FWA leads DeeperForensics-1.0 (0.9334). Xception leads DFD (0.8163). UCF leads DFDC (0.7191), DFDC Preview (0.7594) and UADFV (0.9528). X-ray leads FaceShifter (0.6553). Five different detectors hold the eight column maxima, and the one with the best average, SPSL, leads only two of them.

The within-column spread is as informative as the leader. On Celeb-DF v2 the fifteen detectors sit between 0.6091 and 0.7650, a range of 0.16; on DFD between 0.5481 and 0.8163, a range of 0.27; on DeeperForensics-1.0 between 0.5531 and 0.9334, a range of 0.38. A detector's rank therefore depends heavily on which column is read, which is why the ranked list of detectors on this site uses the average and prints the per-column leaders beside it rather than instead of it.

05How should a row be read?

A row is one detector's profile across eight kinds of unseen data. Xception's row, the most quoted on this site, runs 0.7794, 0.7365, 0.8341, 0.8163, 0.7077, 0.7374, 0.6249, 0.9379: strong on UADFV and DFD, weak on FaceShifter, average 0.7718. EfficientB4 has the same average to four decimals with a different shape: higher on both Celeb-DF columns and UADFV, lower on DFD, DFDC and FaceShifter. Two detectors can tie on the average and still be the wrong choice for a given kind of material.

Two rows show what the grid cannot. X-ray, built to find blending boundaries, scores 0.9391 on its within-domain average but 0.5531 on DeeperForensics-1.0, the lowest DF-1.0 cell by 0.25, because that set's blending leaves no boundary of the kind X-ray learned. Meso4, a 2018 architecture, scores 0.9113 on DeeperForensics-1.0, above Xception, while averaging 0.6551 overall. The grid records that each detector fits some unseen data and not others; it does not say why, and the paper's authors caution that gaps between methods "could be a product of these additional factors rather than the intrinsic superiority of the method". A companion page measures each row against its own within-domain score.

06Why do these numbers differ from other published tables?

Three sources of difference recur. DeepfakeBench's README reproduces Table 3 and notes that the v2 benchmark has run more evaluations and will update the table; this page uses the paper's v2 PDF of 28 October 2023, not the README, and dates every cell 2023. The paper's prose gives UCF's within-domain average as 95.37 percent where the table prints 0.9527; the cross-domain cells on this page have no such discrepancy, but the within-domain figure appears here as the table prints it. Finally, a detector's own paper usually reports higher cross-dataset numbers than this grid, because its authors chose the preprocessing, frame count and augmentation; the benchmark fixes all three for every row, which is the point of a benchmark and the reason the benchmark definition page uses this table as its deepfake example.

Column counts, like which set is hardest, depend on the eight sets chosen. Adding the newer sets catalogued on the dataset timeline would change them; DF40's authors report that on their 40 techniques "existing well-designed (SoTA) detectors may not have obvious advantages over the baseline" (arXiv 2406.13495). No single benchmark score predicts how a detector performs on unseen manipulations, and none of these cells describes an operational system; the TSA checkpoint face comparison page shows what a deployed program publishes instead of a grid like this one.