Deepfake detection benchmarks: datasets, metrics and published results

Public deepfake datasets compared sets the public datasets side by side, with year, manipulation methods, size, source of faces, licence and access route in one table.

FaceForensics++ explains its four manipulation methods, the compression levels its clips come in, and how to request access.

Celeb-DF describes what Celeb-DF set out to fix in earlier datasets, what its test split actually holds, and how to request it.

DFDC covers the challenge dataset: its paid actors, splits and licence, and how the competition was scored.

DeepfakeBench explains the shared codebase that trains and scores many detectors under one protocol, and how to read its tables.

Published detector results collects in-dataset, cross-dataset and in-the-wild scores from papers, each with its train set, test set, protocol, metric and the date we checked it.

How detectors are scored explains AUC, equal error rate, accuracy and false positive rate, and when each one misleads.

Cross-dataset generalisation shows why scores drop on fakes a detector was not trained on, with the published drops model by model.

In the wild looks at what changes when test clips come from real platforms rather than a curated set, with the published drops from WildDeepfake, Deepfake-Eval-2024 and the audio equivalent.

Check a face video yourself sets out the steps a person can take on one clip, ranked by how much evidence each one carries, with the human and detector studies that justify the order.

Can people spot AI faces? reviews what the "which face is real" game and the peer-reviewed human studies measured, how far above chance people got, and what training did.

AI-generated faces explains how synthetic face images are made by adversarial and diffusion generators, what detectors look for in them, and which datasets and results cover them.

Kinds of face deepfake tells face swap, reenactment, attribute editing and full synthesis apart, since each leaves different traces and each dataset covers a different mix.

Liveness and injection attacks covers how identity checks are fooled at the camera or behind it, how presentation attack detection is tested under ISO/IEC 30107-3, and where a deepfake detector fits.

How to read a detection paper walks through where the train set, test set, protocol and metric are stated in a paper, what to do when one is missing, and how a row gets into this site's table.

Evaluate a detector gives a five-step protocol for testing a detector on your own data so the result can be compared with published rows: held-out set, fixed operating point, matched protocol, raw scores, conditions reported.

From face recognition to deepfake detection follows one pipeline from "who is this?" to "is this real?", and shows which benchmark habits carry across and which do not.

Face swap deepfakes traces face swapping from the graphics and autoencoder methods in FaceForensics++ to FaceShifter and SimSwap, and shows why detectors learn the blending boundary.

Deepfake laws lists the deepfake laws of the US, EU, UK and China law by law, with the content each covers, the actor it binds, the platform duty and the penalty, from the official texts.

How accurate are detectors answers the accuracy question with three sourced ranges, for fakes a detector trained on, fakes from a dataset it never saw, and fakes that circulated online in 2024.

Generated-image benchmarks compares the benchmarks that pair real photographs with named generators, from the 2020 ProGAN set to GenImage and the 2024 circulated images, with each table's generator list.

AI image detection explains in plain language what an AI image detector learns, why it generalised across GAN generators in 2020 and lost 45 percent of its AUC on 2024 images, and how to read its score.

Detection glossary defines forty terms used across the detection and face recognition pages in one sentence each, with the paper or standard that defines the term named in the same row.

APCER, BPCER and ACER defines the three presentation attack detection error rates, shows how a result table reports them per protocol, and explains why the standard prefers BPCER at a fixed APCER.

Evaluation protocols separates the three ways a detector is tested, says which question each answers, and follows one detector through all of them with the published AUC.

Human detection studies puts the five published human detection studies side by side, with participants, stimuli, task design, chance level and measured accuracy from each paper.

Papers reading list lists 25 primary deepfake detection papers in reading order, each with its venue, the dataset it rests on and the one result worth quoting.

EU AI Act labelling sets out what Article 50 of the EU AI Act says about marking and disclosing deepfakes, who each duty falls on, the stated exceptions and the application date.

Frame vs video AUC explains why one detector carries a frame-level AUC in one paper and a video-level AUC in another, with the aggregation step that separates them and the published cases where the two were measured.

Deepfake AI, both sides sets the AI that generates deepfakes beside the AI that detects them, and shows that both are trained on recognition networks and scored on the same datasets.

Deepfake images separates swapped, edited and generated still images by the trace each leaves, and reports what image detectors scored on 1,975 fakes that circulated in 2024.

Is this AI generated ranks the kinds of evidence that can answer whether an image is AI generated, from a signed provenance record down to a detector score, with the study behind each.

AI benchmarks explained defines a benchmark as a dataset, a protocol and a metric, and shows saturation with LFW in face recognition and FaceForensics++ in deepfake detection.

Cross-dataset results grid prints every cell of DeepfakeBench Table 3 for the eight unseen test sets, 15 detectors by 8 datasets, and reads the grid by column and by row.

Datasets by year lists the public deepfake detection datasets in release order from 2018 to 2025 with each set's real and fake counts, subjects, methods and footage source.

Detection on social media follows a fake through the platform path, from upload to re-upload, and gives the measured cost of each step to a detector score.

Types of deepfakes puts the four kinds of face deepfake in one table with what each changes, example methods, the datasets that cover each and the detector cue that applies.

Dataset consent lists, for each public deepfake dataset, where its faces came from, what the people shown agreed to, who may download it and how a person can ask to be removed.

DeeperForensics-1.0 describes the 60,000 video DeeperForensics-1.0 set, its 100 consenting actors, its 35 perturbations, its hidden test set and how detectors score on it.

CNN detectors compares the five plain CNN classifiers in DeepfakeBench, from MesoNet to Xception and EfficientNet-B4, on seen and unseen fakes with the published AUC for each.

EfficientNet-B4 detector follows the EfficientNet-B4 row of DeepfakeBench across ten test sets, from 0.9567 on FaceForensics++ c23 to 0.6162 on FaceShifter, and explains why so many detectors build on it.

Deepfake definition in research quotes the sentence each benchmark paper uses to scope the word deepfake, from face swaps only to four kinds of forgery plus audio.

Deepfake laws by state tabulates eight state deepfake statutes from seven legislature websites, with section numbers, scope, penalties and history lines, and says what was not checked.

Face reenactment explains the deepfake kind that keeps identity and transfers expression, from Face2Face in 2016 to the 13 reenactment methods in DF40, with the detector scores each leaves behind.

Passive vs active liveness separates the two liveness approaches by what each reads, what each costs the user and the one attack route both miss, from ENISA, ISO/IEC 30107 and CEN/TS 18099.

ASVspoof 2021 walks through the 2021 edition task by task: trial counts, codecs and transmission conditions, the four baselines, the top five systems per task and the post-challenge results.

Deepfake fraud statistics examines who counts deepfake fraud and how: which official reports isolate it, which do not, what base each figure has and why no total exists.

DFDC log loss explains the metric that ranked the 2020 Kaggle challenge, the public and private test sets behind it, the top five private-set scores and why the paper calls log loss a relative measure.

Unseen generators reads DF40's protocol tables for what happens when a detector meets a generator it was not trained on, including the finding that purpose-built detectors did no better than the baseline.

Detection tips vs evidence lists the common deepfake detection tips in a table and pairs each one with the controlled study that measured whether it helps, with the accuracy figure.

Credentials after upload follows a Content Credential from capture to display, marks the two points where the C2PA manifest can be lost, and quotes what two platforms state about reading and keeping it.

GAN vs diffusion artefacts traces GAN frequency artefacts to upsampling (Frank et al., ICML 2020) and shows what happened when GAN-trained detectors met diffusion images under compression (Corvi et al., 2022).

ForgeryNet documents the ForgeryNet dataset: its clip and image counts, 15 methods in 4 categories, 36 perturbations, four benchmark tasks and the cross-forgery drop its authors measured.

MesoNet traces the 2018 MesoNet networks from their own paper to their rows in the DeepfakeBench table, where Meso4 scores 0.6077 frame-level AUC on FaceForensics++ c23.

Lip-sync deepfakes explains the lip-sync class of deepfake, in which only the mouth is re-rendered to match new audio, with the generators, the datasets that hold them and the mouth-motion detector results.

KYC and deepfakes breaks a remote identity check into capture, liveness, matching and decision, and names the public test behind each step, or the absence of one.

Vision transformers collects the published scores of transformer-based detectors (TALL-Swin, UIA-ViT, CLIP features) and shows which training set, level and protocol sit behind each one.

ASVspoof 5 explains the two tracks of ASVspoof 5, the cost-based metrics that replaced EER, the crowdsourced database and adversarial attacks, and the best submitted scores.

Dataset bias shows, from the Face X-ray and DeepfakeBench tables, that a detector learns the generators and compression of its training set and that the cross-dataset grid is the measurement of that bias.

Detector calibration explains when a detector's score can be read as a probability, how expected calibration error and log loss measure the gap, and why AUC says nothing about it.

What is a deepfake video defines a deepfake video from the FaceForensics++ and DF40 papers, explains frame and clip scoring, and gives the published in-dataset and cross-dataset numbers for the standard detector.

AI or human? Trust reads the third experiment of the 2022 PNAS face study, where synthetic faces were rated more trustworthy than matched real ones, and the typicality explanation its authors give.

AI video watermarks explains what an invisible watermark such as SynthID or a visible AI label proves about a video, what an unmarked clip proves, and how the published robustness figures were measured.

Political deepfakes lists political deepfake cases only where a regulator, court or election authority has described them in a public document, with the date, the medium and the record linked.

UADFV and DeepFake-TIMIT describes the two 2018 sets that started deepfake benchmarking, with their video counts, how the fakes were made, the first detectors tested on them and their role as unseen test sets today.

Face X-ray explains the CVPR 2020 detector that looks for the blending boundary of a pasted face, its training on blended real images only, and the published scores that show where the idea breaks.

Frequency-domain detection explains what the three frequency detectors in DeepfakeBench read in the spectrum, why upsampling leaves a trace there, and what the cross-dataset table says about it.

Certified liveness reads the ISO/IEC 30107-3 testing standard and a NVLAP-accredited lab's published test method to show what a liveness conformance letter states, for which configuration, and what it leaves out.

Deepfake vs AI-generated separates manipulated real footage from fully generated media the way the benchmark papers do, with the counts of each kind in FaceForensics++, DF40 and GenImage.

What makes a deepfake realistic lists the four refinements Celeb-DF made to the basic face swap, resolution, colour, mask and temporal smoothing, and shows how far each generation of fakes pushed detector AUC down.

Reproducing a result walks one published number, Xception at 0.7365 frame-level AUC on Celeb-DF v2, back through the five things a reproducible result has to state, using the DeepfakeBench files that fix each one.

Google DFD dataset describes the 2019 set that Google and Jigsaw recorded with 28 paid, consenting actors and released through FaceForensics++, with its counts, file layout and the published detector scores on it.

Self-blended images explains how the SBI method trains a detector without any real deepfake, lists its published cross-dataset AUC beside FTCN and Face X-ray, and reads it against the DF40 protocol tables.

t-DCF and minDCF explains the tandem detection cost function, its priors and costs, why ASVspoof replaced EER with it for tasks that include a speaker verifier, and what the 2021 and ASVspoof 5 figures mean.

AI video generators lists the text-to-video and image-to-video generators that detection benchmarks have actually tested, with the published F1 and AUC per generator from GenVideo, DF40 and Deepfake-Eval-2024.

Video call deepfakes sets out what primary records say about live face swaps in video calls: the frame rates a swap reaches, the Hong Kong police cases as the Legislative Council recorded them, and the challenge-response test with its AUC.

Video-level detection explains how video detectors score a whole clip from motion rather than single frames, with the video-level AUC that LipForensics, FTCN and TALL report on unseen datasets and the eight video detectors DeepfakeBench v2 lists.

Adversarial attacks reports what two peer-reviewed attack papers measured on named detectors: the AUC left after each attack on an image forensic classifier, and the share of adversarial fake frames that video detectors passed as real, before and after compression.

FaceShifter describes the two-stage FaceShifter swap, its addition to FaceForensics++ for all 1,000 originals in July 2020, and the frame-level AUC each DeepfakeBench detector scores on it.

UCF detector explains the UCF disentanglement detector from ICCV 2023, its 0.9705 AUC on FaceForensics++ c23 in DeepfakeBench, and why its cross-dataset average still trails SPSL.

Deepfake attack vectors separates the three routes a deepfake takes to its target, presentation to a camera, injection behind it and social engineering of a person, with the standard that tests each and the Hong Kong police record.

Multimodal detection explains how joint face and voice detectors are scored on FakeAVCeleb, what single-track models miss, and where the audio-visual numbers come from.

Spoofing vs deepfake detection separates the presentation attack test from the deepfake detection test, with the standard behind each, and shows where both sit in one verification flow.

DFDC Preview describes the October 2019 pilot release of the challenge dataset, its two swap methods, its weighted precision metric and how later benchmarks count and score it.

SPSL detector explains how the CVPR 2021 detector reads the phase spectrum through a shallowed Xception, what its ablation showed, and why it tops the cross-dataset column in DeepfakeBench.

Detection and prevention separates the layers that prevent or record a deepfake from the layer that infers one from pixels, and puts the published figure beside each so the reader knows what each layer is worth.

Xception detector collects every published Xception score, from the FaceForensics++ paper to DeepfakeBench, each with its training set, compression and metric.

AUC score defines the AUC score as a ranking probability, fixes the 0.5 and 1.0 reference points from Fawcett 2006, and shows where published detectors sit on that scale.

How deepfakes work walks the generation pipeline step by step, from face alignment and the shared-encoder autoencoder pair to blending and compression, using the method descriptions in the FaceForensics++, FaceShifter, SimSwap and StyleGAN2 papers.

What is a deepfake defines the term from the FaceForensics++ paper and the dictionary, lists the two families and four types that benchmarks use, and gives the sourced numbers on how well people and detectors tell deepfakes apart.

WildDeepfake describes the dataset of 707 found deepfake videos, how it was collected and labelled, and the accuracy drop every baseline detector showed on it.

Drop by detector lists the in-dataset and cross-dataset AUC of every detector in the DeepfakeBench table and the size of each model's drop.

Deepfake voice explains how text-to-speech and voice conversion produce a cloned voice and reads the ASVspoof 2021 deepfake task results on compressed speech.

Human or AI reads the three controlled studies in which people judged whether a face or a video was human or AI, with the accuracy each reported and what they left out.

Liveness detection defines liveness detection, separates it from deepfake detection and face matching, and explains the ISO/IEC 30107-3 test rules and the OULU-NPU protocols.

How AI faces are made explains how StyleGAN and diffusion models turn random numbers into a face, which photographs they learned from, and what that process leaves behind.

Content Credentials (C2PA) reads the C2PA specification for what a signed manifest proves about a file, what it cannot prove, and what happens when a platform strips it.

Deepfake statistics collects every deepfake figure on this site that traces to a paper or an official report, with what each one measures, and names the repeated numbers that trace to nothing.

Backbones in detectors counts the backbone networks behind the fifteen DeepfakeBench detectors and shows which parts of the face recognition pipeline forgery detection kept.

GenConViT describes the GenConViT video detector, its own reported scores and the figures it posted off the shelf and after fine-tuning on Deepfake-Eval-2024.

Documented incidents lists deepfake incidents only where a government, regulator or court record describes them, with the date, the delivery route and the record linked.

Deepfake meaning traces the word from a method name on online forums to the dictionaries, quotes each first-use date as printed, and maps the everyday sense onto the four kinds that benchmarks use.

DF40 describes the 2024 dataset that generated fakes with 40 techniques on the FaceForensics++ and Celeb-DF footage, its four protocols and eight detectors, and what its tables showed.

ROC AUC draws an ROC curve step by step from a detector's sorted scores, reads operating points, the chance diagonal and the equal error rate off it, and explains why DeepfakeBench reports AUC where FaceForensics++ reported accuracy.

How detection works walks the pipeline from a video clip to a fake or real verdict and sorts published detectors by the cue they learn, with DeepfakeBench scores for each family.

Presentation attack detection explains the four parts of ISO/IEC 30107, how Part 3 counts APCER per attack species and BPCER, the fixed operating point rule, and how a PAD table is read.

Which Face Is Real explains what the two-alternative game built at the University of Washington in 2019 tests, which StyleGAN tells its learn page teaches, why those tells age, and what the controlled studies found.

Why detectors fail separates the three causes of the in-the-wild drop, new generators, platform re-encoding and different real footage, and attaches a published number to each.

Deepfake audio detection explains how speech countermeasures work, how EER and t-DCF score them, and what the ASVspoof 2021 tables and the found-audio studies reported.

Deepfake scams sets out how the recorded deepfake frauds worked, which official bodies have published on them, what their figures measure and what no official number covers.

Provenance vs detection contrasts verifying a signed content history with inferring authenticity from pixels, and explains why the evidence ladder puts the first above the second.

Models ranked ranks all 15 DeepfakeBench detectors by cross-dataset average AUC, with the within-domain score and best unseen set for each.

Deepfake-Eval-2024 describes how the 2024 in-the-wild benchmark was collected and labelled, and what nine open-source detectors, fine-tuned models and commercial models scored on it.

False positive rate defines the false positive rate from the confusion matrix, gives the formula in words, and shows how deepfake detection and face recognition each report it.

Detection techniques compared compares the three detector families DeepfakeBench trains on the same data, with the best in-dataset and cross-dataset average AUC of each family.

Deepfake videos explains what a deepfake video is, how the main benchmarks build and score them, and why the published numbers fall from above 0.97 to below 0.65.

Signs of a deepfake pairs every visible sign of a deepfake with the study that measured how often people actually saw it, by manipulation type and compression level.

Deepfake technology defines deepfake technology by its three model families, paired autoencoders, adversarial networks and diffusion models, and names the datasets each appears in.

ASVspoof lists every ASVspoof edition from 2015 to ASVspoof 5 with its attack types, primary metric, number of teams and the best published error rate.

Compression and detection quantifies what H.264 compression does to detector accuracy on identical fakes, from the FaceForensics++ tables to the DeepfakeBench c40 column and the re-encoded clips of the wild.

Face anti-spoofing covers the datasets, test protocols and error rates of face anti-spoofing, from the four OULU-NPU protocols to the APCER, BPCER and ACER figures papers report.

Best deepfake detection model answers the best-model question from the one shared evaluation, column by column, and explains why the answer changes with the test set.

Equal error rate defines EER as the error at the threshold where misses and false alarms are equal, and shows where face and voice benchmarks report it.

FaceForensics++ c23 and c40 explains the three FaceForensics++ release levels, raw, c23 and c40, and the accuracy drop each one costs the same detector.

Is a deepfake illegal reads the deepfake provisions of US federal, US state, EU and UK law from the official texts, with what each covers and when it took effect.

Detection algorithms glossary lists twenty named detectors in one table with the cue each one reads, its backbone, its venue and its DeepfakeBench cross-dataset average where the paper published one.

Generation methods catalogue walks through the 40 generation techniques the DF40 benchmark catalogued in four families, which ones it held back as unknown domains, and what a FaceForensics++ detector scored on each family.

What is deepfake detection defines deepfake detection as a binary classification task, explains what a detector outputs, how benchmarks score it and why the same model scores differently on different sets.

Humans vs machines on video reports the PNAS 2022 study that put 15,016 people and the leading DFDC model on the same deepfake videos, with every accuracy figure and its condition.

Benchmarks compared puts FaceForensics++, DFDC, DeepfakeBench, DF40 and Deepfake-Eval-2024 in one table by test material, protocol, metric and what each can and cannot show.

Injection attacks explains how a deepfake bypasses the camera in an identity check, why ISO/IEC 30107-3 tests do not score it, and what CEN/TS 18099:2024 and ENISA say about testing it.

Audio deepfakes in the wild reports the Interspeech 2022 test that took twelve audio deepfake detectors trained on ASVspoof 2019 to 37.9 hours of found recordings, model by model.

Accuracy vs AUC explains why a detector with an accuracy of 99 percent on one test set can be useless on another, and what AUC adds and leaves out.