Detecting Deep-Fake Videos from Phoneme-Viseme Mismatches

Recent advances in machine learning and computer graphics have made it easier to convincingly manipulate video and audio. These so-called deep-fake videos range from complete full-face synthesis and replacement (face-swap), to complete mouth and audio synthesis and replacement (lip-sync), and partial word-based audio and mouth synthesis and replacement. Detection of deep fakes with only a small spatial and temporal manipulation is particularly challenging. We describe a technique to detect such manipulated videos by exploiting the fact that the dynamics of the mouth shape - visemes - are occasionally inconsistent with a spoken phoneme. We focus on the visemes associated with words having the sound M (mama), B (baba), or P (papa) in which the mouth must completely close in order to pronounce these phonemes. We observe that this is not the case in many deep-fake videos. Such phonemeviseme mismatches can, therefore, be used to detect even spatially small and temporally localized manipulations. We demonstrate the efficacy and robustness of this approach to detect different types of deep-fake videos, including in-the-wild deep fakes.