Google announced that it is releasing a new large dataset of visual deep fakes as part of the new FaceForensics benchmark. The new dataset, called Deep Fake Detection Dataset is a contribution from Google to the FaceForensics video benchmark which Google is co-sponsoring.
Since 2017 and the time when deep fakes first appeared, the methods for generating fake images and videos of people have advanced quite a lot. Deepfakes have found application in several domains over these years, however many potential dangers of their usage have been identified. For this reason, Google decided to take part in the FaceForensics benchmark and build a dataset of deep fake videos.
The Deep Fake Detection Dataset from Google contains more than 3000 manipulated videos captured in various scenes with 28 actors. In addition, Google also provided the 363 original sequences from 28 paid actors in 16 different scenes, which were modified using Deep Fake methods such as Deepfakes, Face2Face, FaceSwap, and NeuralTextures.
The dataset is available on the following link together with instructions for downloading and loading the data. The original videos are approximately 200GBs while the set of 3000 manipulated videos is more than 1.6TBs.
Google engineers also provided binary masks for all the video manipulation methods. According to them, the users can also download the audio accompanying the videos, which is not part of the dataset.