Best deepfake detection model: what the published tables support

This page belongs to the deepfake detection hub and gives the short answer to a question the deepfake detection models ranked page treats column by column. It names no commercial product: only models with a published paper and a shared evaluation can be compared on equal terms.

01Short answer

Three questions hide inside "which model is best", and the published tables answer each differently. For fakes made the same way as the training set, UCF scores highest: 0.9705 AUC on FaceForensics++ c23 and 0.9527 averaged over the six within-domain test sets in DeepfakeBench Table 3 (arXiv 2307.01426). For fakes made by generators the model never saw, SPSL scores highest on average: 0.7875 across eight cross-domain sets, with the best Celeb-DF v1 and v2 scores of 0.8150 and 0.7650. For fakes that actually circulated online, no model in that table has been scored; the only published in-the-wild figures for open-source video detectors come from Deepfake-Eval-2024, where the best off-the-shelf model, GenConViT, reached 0.63 AUC (arXiv 2503.02857).

Every figure in the table is a frame-level AUC from a detector trained on FaceForensics++ c23, except the Deepfake-Eval-2024 row, which scores models as their authors released them. The results table carries the full set of columns with the same conditions.

02Why is there no single best deepfake detection model?

A detector learns the traces of the generators and the compression in its training set, and the published tables show how far those traces carry. In DeepfakeBench every detector was trained on FaceForensics++ c23 and tested on 14 sets, 6 made from the same source videos and 8 made elsewhere. The within-domain averages sit between 0.93 and 0.96 for the ten strongest models; the cross-domain averages sit between 0.69 and 0.79 for the same models. A model that is best at the first task is not best at the second: UCF tops the within-domain average but SPSL tops the cross-domain average, and Face X-ray, which has the best FaceForensics++ Face2Face score of 0.9872, has one of the weakest cross-domain averages at 0.6985.

The column leaders make the point without an argument. UCF leads nine columns, SPSL three, Face X-ray two, FWA one (DeeperForensics-1.0 at 0.9334) and Xception one (the Google DFD set at 0.8163). No model leads more than nine of sixteen. The cross-dataset generalisation page measures the drop between the two groups of columns for each model.

03Which model is best on fakes like its training data?

UCF. In Table 3 it posts the top within-domain scores on FaceForensics++ c23 (0.9705), c40 (0.8399), the DeepFakes subset (0.9883), the FaceSwap subset (0.9896) and the NeuralTextures subset (0.9441), and the top within-domain average, 0.9527. Its top-3 count within domain is six, the most of any detector. The paper's authors note that the within-domain gap between purpose-built detectors and plain classifiers is small: Xception, a general image classifier with no deepfake-specific design, averages 0.9450 and EfficientNet-B4 0.9389, and the paper states that these "register high AUC values that are comparable to more sophisticated algorithms".

That matters for anyone reading a leaderboard. A within-domain score above 0.94 says that a model has learned the test set's generators, not that it detects deepfakes in general. The DeepfakeBench page describes the protocol that makes these numbers comparable: the same face crop, 32 frames per video, the same augmentations and the average of the top-3 checkpoints.

04Which model is best on unseen fakes?

SPSL, by the cross-domain average, and by a small margin. SPSL averages 0.7875 across the eight sets it never trained on, UCF 0.7801, SRM 0.7760, FFD 0.7733 and Xception 0.7718. SPSL leads both Celeb-DF columns, 0.8150 on v1 and 0.7650 on v2, where UCF scores 0.7793 and 0.7527. UCF leads on DFDC (0.7191 against SPSL's 0.7040), on the DFDC preview (0.7594) and on UADFV (0.9528). The difference between the two leaders on the average is 0.0074 AUC, smaller than the difference between either of them and its own score on a different column.

DF40 (2024), which tested detectors across 40 generation methods, reached a sharper conclusion about the same families of models. Its second finding reads: "Existing well-designed (SoTA) detectors may not have obvious advantages over the baseline", and the authors add that SRM, SPSL, RECCE and RFM, all built on an Xception backbone, "achieve results similar to the simple baseline Xception" (arXiv 2406.13495). Read together, the two benchmarks say the cross-dataset leader changes with the test sets chosen, and the lead is never large.

05Which model is best on fakes from the internet?

None of the DeepfakeBench detectors has a published score on circulated fakes, so the question has to be answered from a different study with different models. Deepfake-Eval-2024 collected 45 hours of video that spread online in 2024 and ran three open-source video detectors on it as released. GenConViT scored 0.63 AUC, down from 0.96 on its own benchmark; FTCN scored 0.50, down from 0.87; Styleflow 0.51, down from 0.95. After fine-tuning on part of the new set, GenConViT reached 0.82, which the paper reports separately because a model that has seen the distribution is no longer off the shelf. The in-the-wild benchmarks page sets these figures beside the audio and image results from the same study.

A reader who needs one model for content found online therefore has a published best of 0.63 AUC off the shelf, and nothing in the research tables that is higher under the same condition. That is the quotable caveat for this page: no benchmark score on a curated set predicts how a detector performs on manipulations it has not seen.

06How to read a "best model" claim

Four conditions turn a score into a comparable one, and a claim that omits any of them cannot be placed in the tables above. The training set, since every DeepfakeBench score comes from FaceForensics++ c23 training. The test set, since the same UCF scores 0.9705 on c23 and 0.7191 on DFDC. The metric and its level, since DeepfakeBench reports frame-level AUC and the FaceForensics++ paper reports video accuracy; the equal error rate page covers a third metric that audio benchmarks prefer. And the compression, since the FaceForensics++ compression levels page shows the same detector losing nearly 15 points of accuracy between light and heavy compression of the same fakes.

The legal question sits beside the technical one. A detector score is evidence about a file; whether making or sharing the file was lawful depends on the use and the place, which the is a deepfake illegal page sets out from the statutes.