What is a deepfake? Definition, kinds, how they are made and measured

This page defines the term as benchmarks use it and is the entry point for the glossary in the deepfake detection hub. The mechanics are on the how deepfakes are made page, and the category by category breakdown is on the kinds of face deepfake page; this one fixes the definition, the taxonomy and the headline numbers.

01What does the word deepfake mean?

The dictionary definition is "a digital image, video, voice, etc. that has been convincingly generated or altered to misrepresent someone as doing or saying something that was not actually done or said", with the etymology "deep (as in deep learning) + fake" and a first known use in 2018 (Merriam-Webster). The research use is narrower. The FaceForensics++ authors write that "The term Deepfakes has widely become a synonym for face replacement based on deep learning", and add that it is also "the name of a specific manipulation method that was spread via online forums" (Rössler et al., 2019). That method, written DeepFakes with a capital F in the paper, is one of the four manipulations in their dataset.

Two things follow for anyone reading a benchmark. "Deepfake" can mean the whole class of learned face and voice forgeries, or one specific autoencoder method, and papers use both senses. And a deepfake is defined by how it was made, not by how convincing it is: a crude swap produced by a trained model is a deepfake, and a convincing cut made by hand in an editing suite is not.

02What kinds of deepfake are there?

The FaceForensics++ paper gives the first working taxonomy: "Current facial manipulation methods can be separated into two categories: facial expression manipulation and facial identity manipulation". Expression manipulation, also called reenactment, keeps the person and changes what the face does; identity manipulation, "known as face swapping", replaces the face with another person's. The dataset holds two methods of each kind.

The DF40 benchmark (2024) widens this to four types: "face-swapping (FS), face-reenactment (FR), entire face synthesis (EFS), and face editing (FE)" (Yan et al., 2024). Entire face synthesis covers faces generated from nothing, as a StyleGAN or diffusion model produces them; face editing changes an attribute such as age while keeping the person. DF40 implements 40 techniques across the four: 10 face swapping, 13 reenactment, 12 synthesis and 5 editing.

Voice is in the dictionary definition and in the research. The ASVspoof 2021 challenge ran a deepfake speech task with 54 teams, in which the top system reached an equal error rate of 15.64 percent, and a 2022 study collected 37.9 hours of found recordings of public figures, 17.2 hours of them deepfakes (Müller et al., 2022).

03How is a deepfake made?

A face deepfake is produced in five steps: gather source and target footage, detect and align the faces, train or load a model, generate a new face for each frame, and blend it back before the video is encoded. In the original DeepFakes method the model is "two autoencoders with a shared encoder", "the trained encoder and decoder of the source face are applied to the target face", and the output is "blended with the rest of the image using Poisson image editing". FaceShifter and SimSwap replaced the per-pair training with an identity encoder borrowed from face recognition, so one source photograph is enough. Generated faces come from a different route: StyleGAN2 produces a 1024 by 1024 face from a random code after training on 70,000 FFHQ photographs (Karras et al., 2020).

The last step is the one that most affects detection. FaceForensics++ releases its clips raw and at two H.264 compression levels, c23 and c40, because "raw videos are rarely found" online.

04How well can people spot a deepfake?

Not well, and worse as quality drops. In the FaceForensics++ user study, 204 participants classified manipulated and real clips correctly 68.69 percent of the time on raw video, 66.57 percent at c23 and 58.73 percent at c40, against a 50 percent chance line. For still faces the result is starker: in Nightingale and Farid's 2022 study in PNAS, 315 participants told StyleGAN2 faces from real photographs with 48.2 percent accuracy, below chance, and a second group given training and feedback reached 59.0 percent (Nightingale and Farid, 2022). The same study found the synthetic faces were rated more trustworthy, 4.82 against 4.48 on a seven point scale.

05How is a deepfake detected?

A detector is a classifier trained on real and manipulated faces that outputs a score per frame or clip. On the FaceForensics++ test split, an XceptionNet classifier trained on all four methods scored 99.26 percent accuracy on raw video, 95.73 percent at c23 and 81.00 percent at c40, so on the dataset it was trained on it beats the human observers at every compression level. Benchmarks report this ability as an AUC rather than an accuracy, because accuracy depends on the threshold and the mix of real and fake in the test set; the AUC score explained page gives the definition and the 0.5 and 1.0 reference points.

The number that defines the field is what happens off the training set. In the DeepfakeBench evaluation (NeurIPS 2023), the same Xception architecture trained on FaceForensics++ c23 reaches a frame-level AUC of 0.9637 on that dataset and 0.7365 on Celeb-DF v2 without retraining (Yan et al., 2023). On fakes collected from the internet in 2024, open-source detectors' AUC fell by about 50 percent for video compared with their original benchmarks, and the best off-the-shelf open-source video model scored 0.63 (Chandra et al., 2025). No single benchmark score predicts how a detector performs on unseen manipulations.

06How are deepfakes benchmarked?

A benchmark is a dataset with a fixed split plus a protocol. FaceForensics++ holds 1,000 real videos and 4,000 manipulated ones; Celeb-DF v2 holds 590 real YouTube videos and 5,639 swaps of celebrities; the Deepfake Detection Challenge set holds over 100,000 clips recorded with 3,426 paid actors who consented to having their likeness modified; WildDeepfake holds 7,314 face sequences cut from 707 deepfake videos found online. The Face Benchmark dataset table lists each with its counts, methods, licence and source of faces, and the WildDeepfake dataset page shows why a set built from found fakes gives lower scores than one built in the lab: image detectors above 95 percent on FaceForensics++ scored below 70 percent on it.

Protocols differ as much as datasets. DeepfakeBench trains every detector on FaceForensics++ c23 and tests on eight other sets, scoring 32 frames per video; DF40 runs four protocols over 40 techniques and reports more than 2,000 evaluations. A score quoted without its dataset, split and protocol is not a result.

07Why do papers define deepfake differently?

Because each dataset fixes the scope its authors needed. The WildDeepfake paper opens with "Deepfake (or "AI faceswap") refers to the set of deep learning-based facial forgery techniques that can swap one person's face in a video to another person", a swap-only definition that matches what its authors could find online in 2020 (Zi et al., 2020). FaceForensics++ takes swaps and reenactments. DF40 takes all four types and motivates the widening by noting that FaceForensics++ "contains out-of-date forgery techniques from the past four years". A detector's published score only covers the kinds of deepfake its test set contains, so the definition a paper uses is part of its result.