You’re not acting like yourself: Deepfake detection based on facial behavior

Politicians and government leaders are critical targets for deepfake attacks. A single deepfake involving these individuals can severely damage their careers or, in extreme

cases, pose a national security threat. Attackers can leverage vast amounts of publicly available audio and video recordings to train their models, making this threat even more pressing. In response, specialized deepfake detectors have been developed to focus on detecting deepfakes targeting a specific Person of Interest (POI). By learning facial expressions and movements unique to the POI, these detectors can identify inconsistencies in deepfakes where these authentic attributes are absent. However,

previous methods relied on Facial Action Units, which offer an incomplete representation of the POI’s behavior. In this paper, we propose a novel approach to learning POI-specific movements without requiring deepfake samples during training, making it independent of any deepfake generation methods. Although our technique is speaker-dependent, it provides a robust solution for protecting high-profile individuals who are particularly exposed to deepfake threats.