Testing human ability to detect ‘deepfake’ images of human faces

Abstract ‘Deepfakes’ are computationally created entities that falsely represent reality. They can take image, video, and audio modalities, pose a threat to many areas of systems societies, comprising topic interest various aspects cybersecurity cybersafety. In 2020, workshop consulting AI experts from academia, policing, government, the private sector, state security agencies ranked deepfakes as most serious threat. These noted since fake material propagate through uncontrolled routes, changes in citizen behaviour may be only effective defence. This study aims assess human ability identify image faces (these being uncurated output StyleGAN2 algorithm trained on FFHQ dataset) pool non-deepfake images random selection dataset), effectiveness some simple interventions intended improve detection accuracy. Using an online survey, participants (N = 280) were randomly allocated one four groups: control group, three assistance interventions. Each participant was shown sequence 20 selected 50 deepfake real faces. Participants asked whether each AI-generated or not, report their confidence, describe reasoning behind response. Overall accuracy just above chance none significantly improved this. Of equal concern fact participants’ confidence answers high unrelated Assessing results per-image basis reveals consistently found certain easy label correctly difficult, but reported similarly regardless image. Thus, although 62% overall, this across ranged quite evenly between 85 30%, with below 50% for every five images. We interpret findings suggesting there is need urgent call action address