Deepfake Videos in the Wild: Analysis and Detection Jiameng Pu† Neal Mangaokar† Lauren Kelly Virginia Tech University of Michigan Virginia Tech [email protected] [email protected] [email protected] Parantapa Bhattacharya Kavya Sundaram Mobin Javed University of Virginia Virginia Tech LUMS Pakistan [email protected] [email protected] [email protected] Bolun Wang Bimal Viswanath Facebook Virginia Tech [email protected] [email protected] ABSTRACT 25, 29, 34, 40, 42, 52] to evaluate the proposed detection schemes. AI-manipulated videos, commonly known as deepfakes, are an In addition, governments are also realizing the seriousness of this emerging problem. Recently, researchers in academia and indus- issue. In the USA, the first federal legislation on deepfakes was try have contributed several (self-created) benchmark deepfake signed into law in December 2019 [14], with the government en- datasets, and deepfake detection algorithms. However, little effort couraging research on deepfake detection technologies. However, has gone towards understanding deepfake videos in the wild, lead- most existing research efforts from academia and industry have ing to a limited understanding of the real-world applicability of been conducted with limited or no knowledge of actual deepfake research contributions in this space. Even if detection schemes videos in the wild. Therefore, we have a limited understanding of the are shown to perform well on existing datasets, it is unclear how real-world applicability of existing research contributions in this well the methods generalize to real-world deepfakes. To bridge this space. A number of questions can be raised: (1) How are deepfake gap in knowledge, we make the following contributions: First, we videos created in the wild? (2) Are deepfakes in the wild different collect and present the largest dataset of deepfake videos in the from deepfake videos produced by the research community? (3) Are wild, containing 1,869 videos from YouTube and Bilibili, and extract deepfakes increasingly appearing in the wild? Are they being viewed over 4.8M frames of content. Second, we present a comprehensive by large populations? (4) Can existing deepfake detection schemes analysis of the growth patterns, popularity, creators, manipulation (primarily tested on deepfakes produced by the research community) strategies, and production methods of deepfake content in the real- accurately detect deepfake videos in the wild? world. Third, we systematically evaluate existing defenses using our This work aims to answer these questions by conducting a new dataset, and observe that they are not ready for deployment in large-scale measurement study of deepfake videos in the wild or the real-world. Fourth, we explore the potential for transfer learning deepfake videos produced and shared by the Internet community schemes and competition-winning techniques to improve defenses. (i.e., not researchers). To the best of our knowledge, this is the largest measurement study to date. Our contributions include the KEYWORDS following: Deepfake Videos, Deepfake Detection, Deepfake Datasets. • We introduce a new deepfake dataset called DF-W, comprised of deepfake videos created and shared by the Internet community. 1 INTRODUCTION We prepare this dataset by scanning a variety of sources on the Web, including YouTube, Bilibili [1], and Reddit.com. Our Advances in deep neural networks (DNNs) have enabled new ways arXiv:2103.04263v2 [cs.CR] 11 Mar 2021 dataset includes a total of 1, 869 videos from YouTube and Bilibili, of manipulating video content. This has fueled the rise of deepfakes, comprising of over 48 hours of video, covering a wide range of or videos where the face of a person is swapped in by another video resolutions. To the best of our knowledge, DF-W is the face, using DNN-based methods. Popular DNN methods for creat- largest collection of deepfake videos in the wild. ing deepfakes include Autoencoders [53], Generative Adversarial • We present a comprehensive analysis of the videos in DF-W. We Networks (GANs) [26], and Variational Autoencoders (VAEs) [33]. examine the differences in content between deepfake videos in Such capabilities to produce convincing deepfakes raises serious DF-W, and datasets released by the research community [12, 25, ethical issues, because they can be misused in many ways, e.g., to 34, 40, 42, 52]. We observe that DF-W videos tend to be more show a person at a place they never went, make a person perform sophisticated, and include several variations of deepfake content, actions they never did, or say things they never said. Such fake thus raising new challenges for detection schemes. We find that videos can help spread fake news, manipulate elections, or incite many DF-W videos are created using generation methods differ- hatred towards minorities [21]. ent from those used by the research community, which poten- Given the potential for misuse of this technology, researchers tially results in a data distribution gap between existing deepfake have proposed a variety of deepfake video detection schemes [20, 21, 27, 40, 48, 49, 59, 61], and also released new deepfake datasets [12, y indicates equal contribution. datasets and DF-W deepfakes. We also analyze the growth, and popularity of deepfake videos in the wild, and investigate the content creators involved in the process. • We systematically evaluate the performance of state-of-the-art deepfake detection schemes on videos in DF-W. We evaluate 7 deepfake detection schemes, including 5 supervised and 2 unsu- pervised schemes. We find that all detection schemes perform poorly on DF-W videos, with the best approach (CapsuleForen- sics, a supervised approach) having an F1 score of only 77% in Figure 1: Frame samples of deepfake videos in the wild catching deepfakes. This means that these existing detection (YouTube). schemes are not ready for real-world deployment. Poor perfor- mance can be attributed to distributional differences between while only using publicly available data and the Tensorflow ma- real-world deepfake videos, and those used to train existing de- chine learning library. It was later revealed that the inspiration for tection schemes. Failure cases can also be partially attributed to their face-swap algorithm was an unsupervised image-translation racial bias, a well known problem with DNN-based facial analy- work from NVIDIA [43]. Since then, Internet communities have sis [51]. We also attempt to interpret the classification decisions produced several deepfake generation tools (see Section 2.2) by using a state-of-the-art model interpretation scheme, called In- leveraging state-of-the-art deep generative models proposed by tegrated Gradients [56]. We leverage this tool to infer features the research community. Such models include Autoencoders [53], that can be used to either improve detection schemes, or create Variational Autoencoders [33], Convolutional Networks [37], and more evasive deepfakes. Generative Adversarial Networks (GANs) [26]. At the same time, • We explore approaches to improve detection performance. Finally, to the research community itself independently produced several vari- improve detection performance on DF-W, we leverage a transfer ants of deepfake generation schemes (see Section 2.2). In this paper, learning-based domain adaptation scheme, which shows promis- we primarily focus on deepfakes produced by the Internet commu- ing results on the DF-W dataset. However, domain adaptation nity, and appearing in the wild. still requires a small number of deepfake videos from the target distribution/domain (DF-W in this case). Therefore, the attacker 2.2 Methods for Generating Deepfakes still has an upper hand, putting the defender in a difficult situa- We describe deepfake generation methods developed by both the tion, unless we come up with defenses that can generalize better. Internet and research communities. To aid our discussion, we adopt We also investigate the performance of the winning DNN model the following convention: the source face is the face to be swapped from Facebook’s DFDC competition [16]. While this winning in, and the target face is the face that will be replaced. model outperforms the existing models (without domain adap- Methods by the Internet community. The following methods tation), its performance on DF-W is still inadequate with an F1 developed by the Internet community, have been used to produce score of 81% and low precision of 71%. deepfakes we find on the Internet. We release the DF-W dataset with the goal of enabling further (1) FaceSwap2: FaceSwap is an open-source tool created in 2017, work on deepfake detection. The DF-W dataset is available on our using two deep autoencoders that share the encoder module, but GitHub repository1. use different decoders. The two autoencoders are trained separately on the source and target faces to reconstruct each face from a latent 2 BACKGROUND representation. The use of the shared encoder enforces a shared latent space, ensuring that the encoder disentangles facial identity 2.1 DeepFake Videos from facial expression. To swap a face, the target face is fed into A deepfake video is popularly characterized as a video that has been the shared encoder, and the source’s decoder is used to decode manipulated using deep neural networks (DNNs), with the goal of the latent representation. The result is the swapped face, i.e., with simulating false visual appearances [23, 27, 34, 41]. We focus on the the source’s face, but the target’s facial expression and pose. The most popular type of deepfakes found on the Internet, categorized as resulting face can then be spliced into the target image. face-swapped videos [21]. This technique attempts to replace the face (2) DeepFaceLab (DFL)3: DFL is another open-source project cre- of an individual with that of another, while retaining the expression, ated in 2018 as a fork of FaceSwap. On their website, it is claimed pose, and background area of the image. Figure 1 shows examples that “More than 95% of deepfake videos are created with DFL”.