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

View on arXiv ← Back to list

Authors: Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, Siwei Lyu

Published: 2019-09-27 21:26:34+00:00

AI Summary

This paper introduces Celeb-DF, a large-scale dataset containing high-quality deepfake videos of celebrities, addressing the limitations of existing datasets with low visual quality. The authors conduct a comprehensive evaluation of deepfake detection methods on Celeb-DF, demonstrating its increased challenge level.

Abstract

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 large-scale datasets. 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.

Key findings

Celeb-DF proves significantly more challenging for existing deepfake detection methods than previous datasets, achieving lower AUC scores across the board. The results highlight the need for improved deepfake detection techniques, especially those robust to high-quality fakes and video compression artifacts.

Approach

The paper focuses on creating a new, high-quality deepfake video dataset (Celeb-DF) to benchmark deepfake detection algorithms. The dataset's high quality stems from improvements in the deepfake synthesis process, addressing issues like low resolution, color mismatch, and inaccurate face masks. The authors then evaluate several existing deepfake detection methods on this new dataset.

Datasets

Celeb-DF, UADFV, DeepFake-TIMIT, FaceForensics++, Google DeepFake detection dataset (DFD), Facebook DeepFake detection challenge (DFDC) datasets.

Model(s)

Two-stream, MesoNet (Meso4, MesoInception4), HeadPose, FWA, VA (VA-MLP, VA-LogReg), Xception (Xception-raw, Xception-c23, Xception-c40), Multi-task, Capsule, DSP-FWA.

Author countries

USA, China

← Previous