FaceForensics++ is a forensics dataset consisting of 1000 original video sequences that have been manipulated with four automated face manipulation methods: Deepfakes, Face2Face, FaceSwap and NeuralTextures. The data has been sourced from 977 youtube videos and all videos contain a trackable mostly frontal face without occlusions which enables automated tampering methods to generate realistic forgeries. Source: https://github.com/ondyari/FaceForensics Image Source: https://github.com/ondyari/FaceForensics

Celeb-DF is a large-scale challenging dataset for deepfake forensics. It includes 590 original videos collected from YouTube with subjects of different ages, ethnic groups and genders, and 5639 corresponding DeepFake videos. Source: https://github.com/danmohaha/celeb-deepfakeforensics Image Source: https://github.com/danmohaha/celeb-deepfakeforensics

The DFDC (Deepfake Detection Challenge) is a dataset for deepface detection consisting of more than 100,000 videos. The DFDC dataset consists of two versions: - Preview dataset. with 5k videos. Featuring two facial modification algorithms. - Full dataset, with 124k videos. Featuring eight facial modification algorithms Source: The Deepfake Detection Challenge (DFDC) Preview Dataset

FaceForensics is a video dataset consisting of more than 500,000 frames containing faces from 1004 videos that can be used to study image or video forgeries. All videos are downloaded from Youtube and are cut down to short continuous clips that contain mostly frontal faces. This dataset has two versions: * Source-to-Target: where the authors reenact over 1000 videos with new facial expressions extracted from other videos, which e.g. can be used to train a classifier to detect fake images or videos. * Selfreenactment: where the authors use Face2Face to reenact the facial expressions of videos with their own facial expressions as input to get pairs of videos, which e.g. can be used to train supervised generative refinement models.

FakeAVCeleb is a novel Audio-Video Deepfake dataset that not only contains deepfake videos but respective synthesized cloned audios as well. Image source: https://arxiv.org/pdf/2108.05080v1.pdf

WildDeepfake is a dataset for real-world deepfakes detection which consists of 7,314 face sequences extracted from 707 deepfake videos that are collected completely from the internet. WildDeepfake is a small dataset that can be used, in addition to existing datasets, to develop more effective detectors against real-world deepfakes. Source: https://github.com/deepfakeinthewild/deepfake-in-the-wild Image Source: https://github.com/deepfakeinthewild/deepfake-in-the-wild

WaveFake is a dataset for audio deepfake detection. The dataset consists of a large-scale dataset of over 100K generated audio clips.

111

DeeperForensics-1.0 represents the largest face forgery detection dataset by far, with 60,000 videos constituted by a total of 17.6 million frames, 10 times larger than existing datasets of the same kind. The full dataset includes 48,475 source videos and 11,000 manipulated videos. The source videos are collected on 100 paid and consented actors from 26 countries, and the manipulated videos are generated by a newly proposed many-to-many end-to-end face swapping method, DF-VAE. 7 types of real-world perturbations at 5 intensity levels are employed to ensure a larger scale and higher diversity. Image Source: https://github.com/EndlessSora/DeeperForensics-1.0

The Korean DeepFake Detection Dataset (KoDF) is a large-scale collection of synthesized and real videos focused on Korean subjects, used for the task of deepfake detection. The dataset consists of 62,166 real videos and 175,776 fake videos from 403 subjects. The fake videos are created using 6 different methods: FaceSwap, DeepFaceLab, FSGAN, FOMM, ATFHP and Wav2Lip.

Localized Audio Visual DeepFake Dataset (LAV-DF). Paper: Do You Really Mean That? Content Driven Audio-Visual Deepfake Dataset and Multimodal Method for Temporal Forgery Localization

Forgery Diversity: DF40 comprises 40 distinct deepfake techniques (both representive and SOTA methods are included), facilitating the detection of nowadays' SOTA deepfakes and AIGCs. We provide 10 face-swapping methods, 13 face-reenactment methods, 12 entire face synthesis methods, and 5 face editing. Forgery Realism: DF40 includes realistic deepfake data created by highly popular generation software and methods, e.g., HeyGen, MidJourney, DeepFaceLab, to simulate real-world deepfakes. We even include the just-released DiT, SiT, PixArt-$\alpha$, etc. Forgery Scale: DF40 offers million-level deepfake data scale for both images and videos. Data Alignment: DF40 provides alignment between fake methods and data domains. Most methods (31) are generated under the FF++ and CDF domains. Using our fake data, you can further expand your evaluation (training on FF++ and testing on CDF).

The detection and localization of highly realistic deepfake audio-visual content are challenging even for the most advanced state-of-the-art methods. While most of the research efforts in this domain are focused on detecting high-quality deepfake images and videos, only a few works address the problem of the localization of small segments of audio-visual manipulations embedded in real videos. In this research, we emulate the process of such content generation and propose the AV-Deepfake1M dataset. The dataset contains content-driven (i) video manipulations, (ii) audio manipulations, and (iii) audio-visual manipulations for more than 2K subjects resulting in a total of more than 1M videos. The paper provides a thorough description of the proposed data generation pipeline accompanied by a rigorous analysis of the quality of the generated data. The comprehensive benchmark of the proposed dataset utilizing state-of-the-art deepfake detection and localization methods indicates a significant drop in performance compared to previous datasets. The proposed dataset will play a vital role in building the next-generation deepfake localization methods. The dataset and associated code are available at https://github.com/ControlNet/AV-Deepfake1M.

Benchmarking initiatives support the meaningful comparison of competing solutions to prominent problems in speech and language processing. Successive benchmarking evaluations typically reflect a progressive evolution from ideal lab conditions towards to those encountered in the wild. ASVspoof, the spoofing and deepfake detection initiative and challenge series, has followed the same trend. This article provides a summary of the ASVspoof 2021 challenge and the results of 54 participating teams that submitted to the evaluation phase. For the logical access (LA) task, results indicate that countermeasures are robust to newly introduced encoding and transmission effects. Results for the physical access (PA) task indicate the potential to detect replay attacks in real, as opposed to simulated physical spaces, but a lack of robustness to variations between simulated and real acoustic environments. The Deepfake (DF) task, new to the 2021 edition, targets solutions to the detection of manipulated, compressed speech data posted online. While detection solutions offer some resilience to compression effects, they lack generalization across different source datasets. In addition to a summary of the top-performing systems for each task, new analyses of influential data factors and results for hidden data subsets, the article includes a review of post-challenge results, an outline of the principal challenge limitations and a road-map for the future of ASVspoof.

We construct the ForgeryNet dataset, an extremely large face forgery dataset with unified annotations in image- and video-level data across four tasks: 1) Image Forgery Classification, including two-way (real / fake), three-way (real / fake with identity-replaced forgery approaches / fake with identity-remained forgery approaches), and n-way (real and 15 respective forgery approaches) classification. 2) Spatial Forgery Localization, which segments the manipulated area of fake images compared to their corresponding source real images. 3) Video Forgery Classification, which re-defines the video-level forgery classification with manipulated frames in random positions. This task is important because attackers in real world are free to manipulate any target frame. and 4) Temporal Forgery Localization, to localize the temporal segments which are manipulated. ForgeryNet is by far the largest publicly available deep face forgery dataset in terms of data-scale (2.9 million images, 221,247 videos), manipulations (7 image-level approaches, 8 video-level approaches), perturbations (36 independent and more mixed perturbations) and annotations (6.3 million classification labels, 2.9 million manipulated area annotations and 221,247 temporal forgery segment labels). We perform extensive benchmarking and studies of existing face forensics methods and obtain several valuable observations.

The quality of AI-generated images has rapidly increased, leading to concerns of authenticity and trustworthiness. CIFAKE is a dataset that contains 60,000 synthetically-generated images and 60,000 real images (collected from CIFAR-10). Can computer vision techniques be used to detect when an image is real or has been generated by AI? ##Dataset details The dataset contains two classes - REAL and FAKE. For REAL, we collected the images from Krizhevsky & Hinton's CIFAR-10 dataset For the FAKE images, we generated the equivalent of CIFAR-10 with Stable Diffusion version 1.4 There are 100,000 images for training (50k per class) and 20,000 for testing (10k per class) ##References If you use this dataset, you must cite the following sources Krizhevsky, A., & Hinton, G. (2009). Learning multiple layers of features from tiny images. Bird, J.J., Lotfi, A. (2023). CIFAKE: Image Classification and Explainable Identification of AI-Generated Synthetic Images. arXiv preprint arXiv:2303.14126. Real images are from Krizhevsky & Hinton (2009), fake images are from Bird & Lotfi (2023). The Bird & Lotfi study is a preprint currently available on ArXiv and this description will be updated when the paper is published. ##License This dataset is published under the same MIT license as CIFAR-10: Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

The FakeMusicCaps dataset contains total of 27605 10 seconds music tracks corresponding to almost 77 hours, generated using 5 different Text-To-Music (TTM) models. It is designed to be used as a starting dataset for the training and/or evaluation of models for the detection and attribution of synthetic music generated via TTM models. If you use this dataset, please cite our paper: Comanducci, L., Bestagini, P., and Tubaro, S., “FakeMusicCaps: a Dataset for Detection and Attribution of Synthetic Music Generated via Text-to-Music Models”, Art. no. arXiv:2409.10684, 2024. doi:10.48550/arXiv.2409.10684.

DeepFake MNIST+ is a deepfake facial animation dataset. The dataset is generated by a SOTA image animation generator. It includes 10,000 facial animation videos in ten different actions, which can spoof the recent liveness detectors.

We created a new dataset, named DFDM, with 6,450 Deepfake videos generated by different Autoencoder models. Specifically, five Autoencoder models with variations in encoder, decoder, intermediate layer, and input resolution, respectively, have been selected to generate Deepfakes based on the same input. We have observed the visible but subtle visual differences among different Deepfakes, demonstrating the evidence of model attribution artifacts.

Human face Deepfake dataset sampled from large datasets - High Quality Dataset - Diverse Dataset - Challenging Dataset - Large Dataset - Text prompts

SONICS is a large-scale dataset comprising 97,164 songs — 48,090 real songs from YouTube and 49,074 fake songs from Suno & Udio — designed for synthetic song detection (SSD), also known as fake song detection (FSD). It addresses several limitations of existing datasets, such as the lack of end-to-end fake songs, limited diversity in music-lyrics, and insufficient long-duration songs. The average length of the songs in SONICS is 176 seconds, which enables the capture of long-context relationships. Moreover, SONICS provides open access to generated fake songs and is divided into 66,709 songs for training, 26,015 songs for testing, and 4,440 songs for validation. Additionally, the inclusion of song lyrics in SONICS dataset paves the way for future research in this field.

DeePhy is a novel DeepFake Phylogeny dataset consisting of 5040 DeepFake videos generated using three different generation techniques. It is one of the first datasets which incorporates the concept of Deepfake Phylogeny which refers to the idea of generation of DeepFakes using multiple generation techniques in a sequential manner. The dataset can be used for the tasks of (i) DeepFake Detection ,(ii) Model Attribution of DeepFakes and (iii) Prediction of the sequential order of DeepFake techniques employed to create phylogenetic deepfakes. It will facilitate advancements in real-life scenarios of plagiarism detection, forgery detection, and reverse engineering of deepfakes.

We release the dataset for non-commercial research. Submit requests <a href="https://forms.gle/6WPEGNWbYoEe6bte8" target="_blank">here</a>. The dataset consists of three parts: Original Samples, Deepfakes, and the Fidelity Score Dataset. The dataset is accessible via Box.com, which offers high availability worldwide and allows for access tracking. ### Dataset Statistics: Number of folder (or videos): - Original : 675 (673k frames with 3 frame sizes each) - Fake (DFL) : 26,085 (5.0M frames) - Fake (Adaptive - DFL) : 7,215 (1.3M frames) - Fake (FSGAN) : 25,433 (4.6M frames) - Fake (LIA) : 25,662 (5.0M frames) DFL quantized ready-to-use models : 60 Fidelity Score Dataset: - Fake : 49,603 videos (32 frames each, 1.35M frames total) - Real : 816 videos (32 frames each, 22k frames total) Human evaluation dataset: 200 videos ### Original Samples While the paper focuses on 8 challenges, we collected 13 in total, which we share in full. The data is provided as frames extracted from raw HD videos recorded at 60 fps. We downsample the frames to approximately 10 fps for ease of use without any expected loss in generation fidelity. The images inside each zip file are face-only. We provide three convenient sizes: 224 x 224, 512 x 512, and 1024 x 1024 pixels. Each zip file contains folders numbered according to the following challenges/tasks: 0. The camera moves around for about two minutes, capturing the face from different angles while the participant sits still. 1. The participant rotates their head from side to side, then looks up and down. Each side should be held for about 5 seconds, rotating as comfortably as possible, totaling around 25 seconds. 2. The participant covers their eyes with a hand, followed by covering the left half, the right half, and finally, the lower half of the face. 3. The participant puts on the provided sunglasses and then takes them off. 4. The participant wears the provided clear glasses, ensuring they reflect a lamp light shining. After setting up the reflection, recording starts. Finally, the participant removes the glasses. 5. The participant moves a green cloth in front of their face 6. The participant puts on a face mask and counts from 1 to 10 out loud. Then, they remove the facemask. 7. The participant presses a finger against a cheek. 8. The participant sticks out a (small) portion of their tongue. 9. The participant laughs for 10 seconds, then frowns, as if angry, for another 10 seconds. 10. The participant slowly stands up and then sits back down. 11. The participant counts from 1 to 50 with distinct breaks between each utterance. 12. The room light is dimmed, and a flashing light is kept on the face for 10 seconds. TODO: Verify each challenge index above corresponds to that challenge in the zips. Note: A few challenges are missing from a handful of participants. ### Fakes Using 47 participants, we create deepfakes between all 2,209 (47 x 47) combinations (including self deepfake), using three deepfake generators — DeepFaceLab, FSGANv2, and LIA. 1. DFL (DeepFaceLab): Notorious for generating hyper-realistic deepfakes, this pipeline serves as a baseline for in-the-wild deepfake videos. For our study, we trained individual DFL deepfake generators for each participant using their `no challenge' videos. These videos records the participants in a range of frontal angles while they sit naturally, aiming to mimic the kind of data readily accessible online for non-celebrity individuals. Training continued until convergence for approximately around 300,000 iterations. DFL (Adaptive Adversary): We utilized heavily trained celebrity deepfake generators provided by the community as an adaptive adversary. Such models are trained on upwards of 2M iterations. The DFL models trained by us and the ones provided by the community are available under dfm_models folder, including the hyperparameters in states.dat file. The models can be loaded directly into Deepfacelive to generate deepfakes. 2. FSGAN (Face Swapping Generative Adversarial Network): This corresponds to the second version of FSGAN. Similar to LIA, this model is also target-agnostic, hence we utilized a pre-trained model made available by the authors for our study. Access its release at https://github.com/YuvalNirkin/fsgan 3. LIA (Latent Image Animator): This pipeline is a facial reenactment method outlined in. Given its target-agnostic nature—meaning it does not require specific target data during inference—we employed a pre-trained model for our experiments. Access its release at https://github.com/wyhsirius/LIA As a rule of thumb, An imposter outer face and target is the inner face, in case of faceswaps. While Box offers quick zipping of multiple files, we have placed the files into zips (<10 GB each), for convenient access. Each zip under fakes (except Adaptive Adversary) is organized in path : “target/challenge/imposter/.jpg”. For Adaptive Adversary, the order is opposite, i.e., “imposter/challenge/target/.jpg” ### Fidelity Score Dataset (or a Snapshot of the full dataset) The training subset contains the following subselections: - Original and deepfake videos with four challenges, one from each category and ‘no challenge’. - 32 frames per sample, - 35 out of 47 target identities, and - Deepfake videos created using only DFL. The validation dataset consists of the remaining 12 out of 47 target identities, while test one includes the celebrity deepfakes. The dataset is the smallest snapshot of the full dataset. The code for loading the dataset and training a fidelity score function can be accessed at https://github.com/mittalgovind/gotcha-deepfakes. ### Human evaluation We release the instruments we used for human evaluation at https://app.gorilla.sc/openmaterials/693684 (with active preview).

DEEP-VOICE: Real-time Detection of AI-Generated Speech for DeepFake Voice Conversion This dataset contains examples of real human speech, and DeepFake versions of those speeches by using Retrieval-based Voice Conversion. Can machine learning be used to detect when speech is AI-generated? ## Introduction There are growing implications surrounding generative AI in the speech domain that enable voice cloning and real-time voice conversion from one individual to another. This technology poses a significant ethical threat and could lead to breaches of privacy and misrepresentation, thus there is an urgent need for real-time detection of AI-generated speech for DeepFake Voice Conversion. To address the above emerging issues, we are introducing the DEEP-VOICE dataset. DEEP-VOICE is comprised of real human speech from eight well-known figures and their speech converted to one another using Retrieval-based Voice Conversion. For each speech, the accompaniment ("background noise") was removed before conversion using RVC. The original accompaniment is then added back to the DeepFake speech: Overview of the Retrieval-based Voice Conversion process to generate DeepFake speech with Ryan Gosling's speech converted to Margot Robbie. Conversion is run on the extracted vocals before being layered on the original background ambience. (Above: Overview of the Retrieval-based Voice Conversion process to generate DeepFake speech with Ryan Gosling's speech converted to Margot Robbie. Conversion is run on the extracted vocals before being layered on the original background ambience.) ## Dataset There are two forms to the dataset that are made available. First, the raw audio can be found in the "AUDIO" directory. They are arranged within "REAL" and "FAKE" class directories. The audio filenames note which speakers provided the real speech, and which voices they were converted to. For example "Obama-to-Biden" denotes that Barack Obama's speech has been converted to Joe Biden's voice. Second, the extracted features can be found in the "DATASET-balanced.csv" file. This is the data that was used in the below study. The dataset has each feature extracted from one-second windows of audio and are balanced through random sampling. **Note: ** All experimental data is found within the "KAGGLE" directory. The "DEMONSTRATION" directory is used for playing cropped and compressed demos in notebooks due to Kaggle's limitations on file size. A potential use of a successful system could be used for the following: Usage of the real-time system. The end user is notified when the machine learning model has processed the speech audio (e.g. a phone or conference call) and predicted that audio chunks contain AI-generated speech. (Above: Usage of the real-time system. The end user is notified when the machine learning model has processed the speech audio (e.g. a phone or conference call) and predicted that audio chunks contain AI-generated speech.) ## Kaggle The dataset is available on the Kaggle data science platform. The Kaggle page can be found by clicking here: Dataset on Kaggle ## Attribution This dataset was produced from the study "Real-time Detection of AI-Generated Speech for DeepFake Voice Conversion" The preprint can be found on ArXiv by clicking here: Real-time Detection of AI-Generated Speech for DeepFake Voice Conversion ## License This dataset is provided under the MIT License: Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

This dataset was created to test whether it's possible to build a general-purpose detector that can tell real images apart from fake ones generated by convolutional neural networks (CNNs), no matter which model or dataset was used to create the fake images. To do this, the authors collected fake images generated by 11 different CNN-based image generation models. These models represent a wide range of current image synthesis techniques and include: ProGAN StyleGAN BigGAN CycleGAN StarGAN GauGAN DeepFakes Cascaded Refinement Networks (CRN) Implicit Maximum Likelihood Estimation (IMLE) Second-order Attention Super-Resolution (SOAT-SR) Seeing-in-the-Dark (SID) The dataset includes fake images from each of these models and a set of real images, allowing for binary classification (real vs. fake). The study found that a standard image classifier (like a convolutional neural network) trained on fake images from just one generator (ProGAN) was able to detect fake images from other, completely different generators with surprising accuracy. This suggests that many CNN-generated images, even from different architectures, share common flaws that can be learned and detected. The dataset is useful for research in detecting synthetic media, improving image forensics, and understanding the weaknesses in current generative models. Code and pre-trained models were made available by the authors (https://github.com/chuangchuangtan/GANGen-Detection).

The Deepfake face detection task involves a facial image of unknown authenticity for testing. While most deepfake detection methods take only the image as input, our literature demonstrates that conditioning the deepfake detector on identity—i.e., knowing whose deepfake face the picture might be—can enhance detection performance. Existing deepfake detection datasets, such as FaceForensics++ and DFDC, do not include identity information for authentic and deepfake faces. This dataset contains facial images of 45 specific individuals, divided into train and test sets, including a total of 23k authentic and 22k deepfake images. Having a specific individual's images in both the train and test sets allows us to assess detection performance for that individual. The dataset is curated so that the train and test sets are from two independent sources. The train images are curated from the CelebDFv2 dataset, and the test images are curated from the CACD dataset. Deepfake faces are generated using FaceswapGAN, utilizing a portion of the training images to train the reconstruction model. The test deepfake images are faceswapped with another identity not included in our celebrity list. On the other hand, the training deepfake images are reconstructed images of that person. The deepfake detection method proposed in our paper requires reconstructing both the training and test images. The reconstructed test and train images are also available in this dataset. It is worth mentioning that reconstructing the training deepfake images produces doubly reconstructed images.

ShiftySpeech: A Large-Scale Synthetic Speech Dataset with Distribution Shifts ## 🔥 Key Features - 3000+ hours of synthetic speech - Diverse Distribution Shifts: The dataset spans 7 key distribution shifts, including: - 📖 Reading Style - 🎙️ Podcast - 🎥 YouTube - 🗣️ Languages (Three different languages) - 🌎 Demographics (including variations in age, accent, and gender) - Multiple Speech Generation Systems: Includes data synthesized from various TTS models and vocoders. Dataset can be downloaded from: Hugging Face

Synthetic Speech Attribution Dataset.

The TwinSynths dataset is a novel benchmark designed to overcome common limitations found in earlier synthetic image datasets, such as low image quality, inadequate content preservation, and limited class diversity. TwinSynths generates pairs of images where each synthetic image is visually identical to its real counterpart, ensuring that the essential content remains intact while showcasing the unique architectural features of the generative models used. TwinSynths comprises two subsets: TwinSynths-GAN This subset uses a GAN generator architecture which trained from scratch on individual real images using a mean-squared error loss to ensure pixel-level fidelity. By fixing the latent vector input, the method produces synthetic images that closely mirror the original content. The GAN subset consists of 8,000 generated images spanning 80 classes selected from ImageNet. TwinSynths-DM For the diffusion model-based subset, DDIM inversion is used to maintain the content integrity of the original images. By applying a noise-adding forward process followed by a text-conditioned denoising procedure (using class name prompts), this generates synthetic images that are highly similar to their real counterparts. The same set of ImageNet classes is used as in the GAN subset, allowing for a consistent evaluation across different generative methods.

The DeepSpeak dataset contains over 43 hours of real and deepfake footage of people talking and gesturing in front of their webcams. The source data was collected from a diverse set of participants in their natural environments and the deepfakes were generated using state-of-the-art open-source lip-sync and face-swap software.