Deepfake Detector Benchmark Leaderboard: Living Scores From One Standard Test

Every deepfake detector page on the internet leads with an accuracy number, and almost none of those numbers can be compared. Each vendor picks its own test set, its own split and its own scoring rule, so a 99 percent claim and a 91 percent claim usually describe completely different exams. Our accuracy comparison walks through why vendor claims do not line up. This page takes the other route: one public benchmark, one protocol, every detector retrained and retested the same way, ranked in the open.

This is a living leaderboard. Version 1 was published on October 28, 2026, and the scores will be refreshed as new standardized results are published. Tools that have never been measured under this protocol do not appear in the table, and that includes our own DeepfakeDetect app. Unranked is not a judgment on any tool. It is simply what honest bookkeeping looks like: nobody joins the ranking until they have been measured the same way as everyone else.

The protocol behind the ranking

Source: DeepfakeBench (Yan et al., NeurIPS 2023 Datasets and Benchmarks Track), Table 2. All 15 detectors were trained on FaceForensics++ (c23) and evaluated by AUC, with the benchmark reporting the average of its top 3 runs. The cross-domain average is the mean AUC across the 8 unseen test sets: Celeb-DF v1, Celeb-DF v2, DeeperForensics-1.0, DeepFakeDetection, DFDC, DFDC Preview, FaceShifter and UADFV. Read the paper or browse the code and data.

The leaderboard

Ranked by cross-domain average AUC, the column that measures performance on faces, generators and pipelines a detector never saw in training. The FaceForensics++ column is the in-domain score, shown alongside so the drop is visible rather than implied.

Top 8 detectors by cross-domain average AUC. Every detector was trained on FaceForensics++ (c23) and scored by AUC under the DeepfakeBench protocol (Yan et al., NeurIPS 2023), Table 2.

DeepfakeBench Table 2 scores: AUC after training on FaceForensics++ (c23). The cross-domain average covers Celeb-DF v1, Celeb-DF v2, DeeperForensics-1.0, DeepFakeDetection, DFDC, DFDC Preview, FaceShifter and UADFV.
Rank	Detector	Type	FF++ (c23) AUC	Celeb-DF v2 AUC	DFDC AUC	Cross-domain avg AUC
1	SPSL	Frequency	0.9610	0.7650	0.7040	0.7875
2	UCF	Spatial	0.9705	0.7527	0.7191	0.7801
3	SRM	Frequency	0.9576	0.7552	0.6995	0.7760
4	FFD	Spatial	0.9624	0.7435	0.7029	0.7733
5	Xception	Naive	0.9637	0.7365	0.7077	0.7718
6	EfficientNet-B4 (shown as EfficientB4 in the paper)	Naive	0.9567	0.7487	0.6955	0.7718
7	CORE	Spatial	0.9638	0.7428	0.7049	0.7694
8	RECCE	Spatial	0.9621	0.7319	0.7133	0.7649
9	F3Net	Frequency	0.9635	0.7352	0.7021	0.7645
10	Capsule	Spatial	0.8421	0.7472	0.6465	0.7488
11	MesoInception	Naive	0.7583	0.6966	0.6226	0.7364
12	DSP-FWA (shown as FWA in the paper)	Spatial	0.8765	0.6680	0.6132	0.7239
13	CNN-Aug	Naive	0.8493	0.7027	0.6361	0.7020
14	Face X-ray	Spatial	0.9592	0.6786	0.6326	0.6985
15	MesoNet (shown as Meso4 in the paper)	Naive	0.6077	0.6091	0.5560	0.6551

What the leaderboard actually says

In-domain, the leaders are bunched together. The top group scores between 0.957 and 0.971 AUC on FaceForensics++ (c23), and the benchmark authors note that simple naive detectors, Xception and EfficientNet-B4, score comparably to far more sophisticated methods once everyone follows one protocol. On home turf, architecture is not what separates the field.

Cross-domain, everyone falls. SPSL, the leader, averages 0.7875 across the 8 unseen sets, about 17 points below its own in-domain score of 0.9610. DFDC is the hardest of the widely reported tests here: the best any detector manages on it is 0.7191, from UCF. A detector that looks finished in-domain is, on unfamiliar data, closer to a strong signal than a verdict.

Two of the top three cross-domain spots belong to frequency-domain methods: SPSL at number 1 and SRM at number 3, with UCF, a spatial method, between them. Whatever survives a change of dataset appears to live partly in frequency artifacts that spatial features miss.

Face X-ray is the cautionary tale, in both directions. It scores 0.9592 in-domain, close to the top of the table, then lands second from the bottom on cross-domain average at 0.6985. At the same time, the benchmark authors note that blending-based methods stand out in cross-manipulation tests, where the manipulation type changes but the dataset does not. Cross-manipulation and cross-dataset are different exams. A detector can be excellent at one and ordinary at the other, and a single headline number hides which exam was taken.

How to read a vendor accuracy claim

Ask which dataset the number comes from, and whether the detector was trained on that same dataset. In this table, in-domain scores run far above cross-domain scores for the same detector.
Ask whether the score is frame-level or video-level. Aggregating frames into a single video score usually flatters the result.
Ask which generators and manipulations the test set contains. A test built from one manipulation family says little about the next one.
Treat any 99 percent claim as an in-domain result until proven otherwise. Nothing in a standardized cross-domain test comes close to it.

What would move this leaderboard

New standardized suites, applied to every detector under one protocol. DF40 and the continuing DeepfakeBench updates are the candidates we are watching for the next refresh. When their results are published in comparable form, this table gains new rows, the version number moves, and the date at the top of this page changes with it. Until then, version 1 stands as published on October 28, 2026.

If you need a check today rather than a ranking, theDeepfakeDetect app screens images, text, video and audio in the browser, with 3 checks a day. It stays unranked here on purpose, for the reason given above. For the wider commercial field, see the 10 tools worth comparing and our guide to checks for image, text and voice. The moment a same-protocol score exists for any of them, this page is where it lands.