A Case Study on how Beautification Filters Can Fool Deepfake Detectors

Abstract

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.