Deepfake generation methods: the 40 techniques DF40 catalogued and how detectors fared on them

This page is part of the deepfake detection hub and reads the catalogue from the detection side: which methods exist, how the paper groups them and what happens when a detector meets one it was not trained on. The DF40 dataset page covers the dataset itself, and the twenty named detectors page lists the models that are tested against these generators.

01What are the four families of deepfake generation?

DF40 defines each family by what it changes in a target face. Face swapping (FS) replaces the content of the target with that of a source "preserving the identity" of the source, so the person changes and the pose and background stay. Face reenactment (FR) keeps the target's identity "but has its intrinsic attributes, e.g., pose, mouth and expression manipulated" by a driving signal, which may be another video, a set of landmarks, an audio track or text. Entire face synthesis (EFS) "generates a entirely synthesis face" from noise, with no real target at all. Face editing (FE) alters "external attributes, such as facial hair and age" and leaves identity and expression alone (arXiv 2406.13495, section 2.1).

The same four labels organise the kinds of face deepfake page on this site, which explains the visual difference between a swap and a reenactment. Here the labels matter for a different reason: the paper's results are reported per family, and the drop a detector suffers depends on which family it was trained on and which one it is tested on.

02The 40 techniques DF40 catalogued

The table below is DF40's Table 2 reduced to five columns. "Sub-type" is the paper's own tag for the mechanism. "Use" says whether DF40 generated training and test data with the method on FaceForensics++ and Celeb-DF footage, or held it back as a test-only set on other footage. Train-and-test video methods have 1,500 or more clips each and image methods 48,000 or more images; the test-only sets run from 50 clips to 2,000 images.

Row 28 keeps the paper's own tag even though the method is a diffusion model, because the table reproduces Table 2 as printed. The DF40 README adds that 31 of the 40 methods were generated on the FaceForensics++ and Celeb-DF domains and that the whole set is released under a CC BY-NC 4.0 licence by Google form.

03How did earlier datasets catalogue generation methods?

Before DF40 the catalogue was short. FaceForensics++ (ICCV 2019) used four methods, two swaps and two reenactments, with NeuralTextures the newest; the Deepfake Detection Challenge set used eight, five swaps, one reenactment and two synthesis methods; ForgeryNet (CVPR 2021) reached 15 across all four families, according to DF40's comparison table. DF40 states that "most earlier public deepfake datasets published before 2021 generate deepfakes using only the face-swapping technique that involves a blending process", and that they "contain only single or no more than 4 specific manipulation approaches".

That history explains why the paper calls FaceForensics++ a source of "out-of-date forgery techniques from the past four years": a detector trained on it has seen two swaps and two reenactments, and none of the twelve synthesis or five editing methods above. The comparison of public datasets lists each set's method count beside its clip count.

04Does a detector trained on FaceForensics++ generalise to them?

The paper's protocols answer this directly, with Xception as the baseline. Trained on the DF40 face swaps made from FaceForensics++ footage, Xception scored 0.991 frame-level AUC on other swaps from the same footage, 0.892 on reenactments and 0.810 on synthesised faces (Table 3). Moving the test footage to Celeb-DF while keeping the training set fixed gave 0.922, 0.657 and 0.642 (Table 4), which the paper describes as "nearly a 20% drop" for the two unseen families.

On the nine test-only sets that stand in for an unknown domain, the same model averaged 0.546 AUC: 0.882 on DeepFaceLab swaps, 0.394 on HeyGen reenactments, 0.577 on StarGAN edits and 0.616 on StarGANv2 edits (Table 5). The paper's first finding names the pattern: "asymmetric performance drop among different forgery types (fake regions matter)", with swap-trained models holding around 0.8 on synthesis while synthesis-trained models reach only about 0.6 on swaps. Its second finding is that well-designed detectors "may not have obvious advantages over the baseline", and its third that CLIP generalised better than Xception, reaching 0.692 on the unknown sets.

05Which methods did DF40 hold back as unknown domains, and why?

Nine of the forty appear only as test data: DeepFaceLab, HeyGen, the two FFHQ-based synthesis sets, and five editing methods on CelebA. The paper's note states that this data "is utilized as the unknown domain data, simulating the domain shift of real people distributions", and that the two one-to-one methods got only 50 and 100 clips because "their implementation requires significantly more time and resources".

One further finding concerns a method in the table: reenactment forgeries "may share transferable patterns for detecting other FR instances", with Wav2Lip the exception because it "only modifies the mouth region" (Finding 5).

06What this catalogue does not say

DF40 catalogues methods by output and mechanism; it does not rank them by realism, and its authors list the thin coverage of video-level detectors as a limitation. No single benchmark score predicts how a detector performs on a manipulation it has not seen, and the 40-method table above is the clearest published list of what "unseen" can mean. The definition of deepfake detection explains how the scores in this page are computed.