The Celeb-DF-B is a novel dataset created by applying popular social media beautification filters to a subset of real and fake videos from the Celeb-DF dataset. It was created with the objective of evaluating the effect of beautification fitlers on three SotA passive deepfake detectors and on human evaluators.
The Celeb-DF-B dataset is composed of 928 videos balanced in terms of four categories Real, Real-Beautified, Fake, Fake-Beautified.
Motivation
The exponential growth of shared multimedia content made necessary algorithms for deepfake detection. At the same time, beautification filters have become a popular tool and new filters are released every day. Therefore, is it possible to fool state-of-the-art (SotA) detectors by simply applying a beautification filter to the manipulated video?
In order to study the impact of beautification filters on different AI-based deepfake detectors we propose the Celeb-DF-B face dataset. In our work we proved that even easy-to-use social media filters can significantly increase the likelihood of a deepfake video being wrongly classified as authentic.
In addition, we show how the application of beauty filters changes the perceived information for human altering their ability to effectively distinguish deepfake videos.
Source Datasets
The original videos present in the Celeb-DF-B dataset were selected from the publicly available Celeb-DF dataset which consists of 590 real videos and 5639 DeepFake videos with an average duration of 13 seconds per video and a standard frame rate of 30 frames per second.
For the creation of Celeb-DF-B, we chose a subset consisting of 232 real and 232 fake videos of Celef-DF dataset. The selection of videos followed three criteria: 1) an equal sampling from each identity in the real videos; 2) pairing each real video with a fake counterpart created through FaceSwap; and 3) maintaining a balance in the number of time one identity is used as source and target.
Pipeline
A subset of 464 videos (50% Real and 50% Fake) are selected from Celeb-DF dataset. Each video is uploaded to the social network Instagram, where one of the four different filters is randomly selected and applied to it. The four filters uniformly appear in the Celeb-DF-B database. The final database has a size of 928 videos and it is used to perform a human-based deepfake detection and to evaluate the robustness of three SotA AI-based detectors.
Reference
Any publication using this database must cite the following paper:
LIBOUREL, Alexandre, et al. A Case Study on How Beautification Filters Can Fool Deepfake Detectors. En IWBF 2024, 12th IEEE International Workshop on Biometrics and Forensics. 2024.
@inproceedings{libourel2024case, title={A Case Study on How Beautification Filters Can Fool Deepfake Detectors}, author={Libourel, Alexandre and Husseini, Sahar and Mirabet-Herranz, Nelida and Dugelay, Jean-Luc}, booktitle={IWBF 2024, 12th IEEE International Workshop on Biometrics and Forensics}, year={2024} }
