A case study on how beautification filters can fool deepfake detectors

Libourel, Alexandre; Husseini, Sahar; Mirabet-Herranz, Nélida; Dugelay, Jean-Luc
IWBF 2024, 12th IEEE International Workshop on Biometrics and Forensics, 11-12 April 2024, Twente, Netherlands

If telling the difference between a real video and a deepfake is difficult, with the proliferation of beautification filters on social networks it becomes nearly impossible to differentiate between a real video, a video enhanced by a filter, and a video with its original identity replaced. Therefore, is it possible to fool state-of-the-art (SotA) detectors by simply applying a beautification filter to the manipulated video? In this paper, we
study the impact of beautification filters on Celeb-DF-B, a novel database created by applying popular social media beautification filters to a subset of real and fake videos from the Celeb- DF dataset. We assessed three SotA passive deepfake detectors, comparing their performance against that of human evaluators. The results indicate that filters significantly alter the behavior of the three detectors studied, resulting in a notable decrease in the video-level AUC when classifying beautified videos. In the context of human-level performance, the use of filters similarly influences human decision-making, affecting the accurate categorization of videos as either real or fake.

Sécurité numérique
Eurecom Ref:
© 2024 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

PERMALINK : https://www.eurecom.fr/publication/7648