Celeb-DF: A Large-Scale Challenging Dataset for DeepFake Forensics

By Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, Siwei Lyu. 2019

1,748 citations recorded by the source corpus.

AI-synthesized face-swapping videos, commonly known as DeepFakes, is an emerging problem threatening the trustworthiness of online information. The need to develop and evaluate DeepFake detection algorithms calls for datasets of DeepFake videos. However, current DeepFake datasets suffer from low visual quality and do not resemble DeepFake videos circulated on the Internet. We present a new large-scale challenging DeepFake video dataset, Celeb-DF, which contains 5,639 high-quality DeepFake videos of celebrities generated using improved synthesis process. We conduct a comprehensive evaluation of DeepFake detection methods and datasets to demonstrate the escalated level of challenges posed by Celeb-DF.

Subfields: Computer Vision and Pattern Recognition, Evaluation & Benchmarks, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Evaluation & Benchmarks

In Safety, Robustness & Interpretability

Citations in the corpus

CoReD: Generalizing Fake Media Detection with Continual Representation using Distillation (2021, background): “Recently, several more advanced deepfake datasets (DFDC [17], CelebDF [45], CelebFOM [64], and WildDeepfake [80]) as well as GAN image datasets (CelebBlend [64], GANsformer [25], and StytleGAN2 [34]) are released.”
Lips Don't Lie: A Generalisable and Robust Approach to Face Forgery Detection (2020, methodology): “datasets [46, 38, 30, 16], it is possible to train deep convolutional neural networks (CNNs) to detect fake videos [3, 25, 64, 4, 14, 45, 42, 36].”
Media Forensics and DeepFakes: An Overview (2020, background): “Also in [233] a new deepfake video dataset has been introduced,”
Spatial-Phase Shallow Learning: Rethinking Face Forgery Detection in Frequency Domain (2021, methodology): “AUC on FF++. Results for some other methods are from [26], and the bold results are the best. inal XceptionNet method.”
M2TR: Multi-modal Multi-scale Transformers for Deepfake Detection (2021, methodology): “Given that Celeb-DF [37] contains high-quality samples (as discussed in 4.2, Celeb-DF achieves the best results on Mask-SSIM , Perceptual loss and Ewarp metrics in the available Deepfake dataset.)”
Detecting Deepfakes with Self-Blended Images (2022, methodology): “Although the latter methods perform well on some datasets [2, 42], low-quality videos in more challenging datasets [19, 20] where artifacts are hardly recog-”
Learning Self-Consistency for Deepfake Detection (2020, methodology): “To demonstrate the effect of I2G , we conduct ablation studies by training on either DF or DFDC-P and benchmarking on DFR, CD2, DFDC, and DFDC-P test sets.”
One Shot Face Swapping on Megapixels (2021, background): “Besides, it is time consuming and difficult for subject-specific methods to train specific models for distinct pairs of faces [14, 46, 45, 29, 20, 55, 26].”
ContraGAN: Contrastive Learning for Conditional Image Generation (2020, background): “Besides, it would be good to study the general characteristics of generated samples [63] and how we can distinguish fake images from unknown generative models [64, 65, 66].”
Local Relation Learning for Face Forgery Detection (2021, neutral): “Face X-ray (Li et al. 2020a) localizes the forged boundary in a self-supervised mechanism, achieving remarkable performance at high-quality data.”
Video Transformer for Deepfake Detection with Incremental Learning (2021, methodology): “After we extract image features through XceptionNet, we get a feature vector of dimension [2048, 10, 10], we then pass this feature vector to a 2D convolutional layer and a linear layer, which gives us a feature vector of dimension [1, 32, 768].”
FakeCatcher: Detection of Synthetic Portrait Videos using Biological Signals (2019, neutral): “e [37] 96.37% 96.00% 96.75% FF [2] Face2Face [37] 96% 94.24% 97.75% FF++ [3] FaceSwap [59] 95.75% 94.75% 96.75% FF++ [3] Deepfakes [57] 94.87% 93.25% 96.50% FF++ [3] All 94.65% 88.25% 96.25% Celeb-DF [4] Default 91.50% 76.40% 99.41% DF (ours) Mixed 91.07% 85.26% 96.89% FF++ [3] Neural Textures [77] ”
The DeepFake Detection Challenge (DFDC) Dataset. (2020, positive): “Furthermore, as we observed that many publicly-released datasets [14, 17, 18, 23, 30] did not guarantee that their subjects were willing participants or agreed to have their faces modified, we solicited video data from 3,426 paid actors and actresses speaking in a variety of settings for roughly 15 ”
Exploring Temporal Coherence for More General Video Face Forgery Detection (2021, methodology): “To evaluate the generalization capability of our framework, we test our model on the following datasets: 1) FF++ that contains four types of manipulation as mentioned above; 2) FaceShifter[31](FSh) and 3) DeeperForensics[28](DFo) employ the real videos from FF++ for high-fidelity face swapping; 4) D”
The DeepFake Detection Challenge Dataset (2020, methodology): “Furthermore, as we observed that many publicly-released datasets [14, 17, 18, 23, 30] did not guarantee that their subjects were willing participants or agreed to have their faces modified, we solicited video data from 3,426 paid actors and actresses speaking in a variety of settings for roughly 15 ”
FakeAVCeleb: A Novel Audio-Video Multimodal Deepfake Dataset (2021, methodology): “Keeping in mind the essence of datasets, researchers have also proposed different deepfake datasets, generated using the latest deepfake generation methods [9, 11, 15, 19, 22, 32] to help researchers train and evaluate their deepfake detection methods.”