Zoom-DF: A Dataset for Video Conferencing Deepfake

Abstract

With the rapid growth of deep learning methods, AI technologies for generating deepfake videos also have been significantly advanced. Nowadays, the manipulated videos such as deepfakes are so sophisticated that one cannot easily differentiate between real and fake, and one can create such videos with little effort. However, such technologies can be likely to be abused by people with malicious intents. To address this issue, approaches and efforts to detect deepfakes have been researched significantly. However, the performances of the detectors in general depends on the quantity and quality of the training data. In this paper, we introduce a new deepfake dataset, Zoom-DF, which can be injected during the remote meeting and video conferencing, to create a sequence of fake participant images. While most deepfake datasets focus on the face area, our dataset primarily targets for the remote meeting, and manipulates movements of the participants. We evaluate existing deepfake detectors on our new Zoom-DF dataset and present the performance results.