FakeAVCeleb
About
In FakeAVCeleb, we propose a novel Audio-Video Deepfake dataset (FakeAVCeleb) that contains not only deepfake videos but also respective synthesized lip-synced fake audios.
Our FakeAVCeleb is generated using recent most popular deepfake generation methods.
To generate a more realistic dataset, we selected real YouTube videos of celebrities having five racial backgrounds (Caucasian (American), Caucasian (European), African, Asian (East), and Asian (South)) to counter the racial bias issue.
Our works are summarized as follows:
We present a novel Audio-Video Multimodal Deepfake Detection dataset, FakeAVCeleb, which contains both, video and audio deepfakes with accurate lip-sync. Such a multimodal deepfake dataset has not been developed in the past.
Our FakeAVCeleb dataset contains three different types of Audio-Video deepfakes, generated from a carefully selected real YouTube video dataset using recently proposed popular deepfake generation methods.
The individuals in the dataset are selected based on four major ethnic backgrounds speaking the English language to eliminate racial bias issues. Moreover, we performed the comprehensive baseline benchmark evaluation and demonstrated the crucial need and usefulness for a multimodal deepfake dataset.
Samples from each category: