ScholarWorks@성균관대학교

초록

Deepfake harm arises not from technical manipulation alone, but from the moment viewers perceive the synthesized face as a specific victim. Yet the threshold at which this identity perception occurs remains undefined, creating gaps in both detection systems and legal frameworks. We investigate the conditions under which synthesized faces are perceived as target individuals and examine alignment between human perception and computational similarity models. Through a study with 102 participants evaluating 162 manipulated face stimuli, we identify three key findings: (1) eyes serve as the dominant cue for identity perception; (2) a distinct identity threshold emerges when eyes are combined with one additional feature; and (3) human-AI misalignment peaks at this threshold, where humans perceive the victim but AI models do not. These findings highlight the need for perception-aligned detection systems and victim-centered regulatory approaches in addressing deepfake harms.

키워드