Lillie C. del Real¹ (), Jessica N. Goetz¹, Mark B. Neider¹; ¹University of Central Florida

Several studies have demonstrated that human deepfake classification accuracy is often near chance levels and synthetic faces are rated as more trustworthy than real faces (e.g., Nightingale & Farid, 2022). These findings may suggest the potential for subjective judgments like trustworthiness and related attributes, like attractiveness and symmetry, to influence classification accuracy of real and synthetic faces (Bascandziev & Harris, 2014; Perrett et al., 1999). In the current study (N = 418) participants rated (e.g., trustworthiness, attractiveness, symmetry, facial expression, racial ambiguity, threat, competency, masculinity/femininity, likeableness, and unusualness) and classified 160 high-quality face images (80 real; 80 synthetic generated using NVIDIA’s StyleGAN2 (Karras et al., 2020)) as real or synthetic to examine whether a participant’s perceptual experience of a face contributes to how they classify that face. Overall classification accuracy was slightly above chance performance (~52%) (p < .001) with synthetic faces rated more trustworthy, attractive, and symmetrical than real faces (all ps < .001). Two binary logistic regressions of the subjective ratings as predictors of classification accuracy of synthetic faces and real faces revealed aggregate ratings of attractiveness, symmetry, unusualness, and likeableness as positive predictors of correct classification of synthetic faces and negative predictors of correct classification of real faces (all ps ≤ .05). Additionally, to explore more complex patterns in the data, we trained a support vector machine (SVM) using the image ratings as features. The SVM was able to correctly classify a face as real or synthetic with 65% accuracy (relying heavily on attractiveness, unusualness, and symmetry ratings). Combined, our findings suggest that while humans struggle to classify synthetic faces, their percepts of those faces may hold some predictive power with respect to accurate classification and potential for intervention.