Synthetic and natural face identity processing share common mechanisms
Kim Uittenhove, Hatef Otroshi Shahreza, Sébastien Marcel, Meike Ramon
Abstract
Recent developments in generative AI offer the means to create synthetic identities, or deepfakes, at scale. As deepfake faces and voices become indistinguishable from real ones, they are considered as promising alternatives for research and development to enhance fairness and protect humans' rights to privacy. Notwithstanding these efforts and intentions, a basic question remains unanswered: Are natural faces and facial deepfakes perceived and remembered in the same way? Using images created via professional photography on the one hand, and a state-of-the-art generative model on the other, we investigated the most studied process of face cognition: perceptual matching and discrimination of facial identity. Our results demonstrate that identity discrimination of natural and synthetic faces is governed by the same underlying perceptual mechanisms: objective stimulus similarity and observers’ ability level. These findings provide empirical support both for the societal risks associated with deepfakes, while also underscoring the utility of synthetic identities for research and development. • Study examines perception (matching/discrimination) of natural faces vs. facial deepfakes. • Natural identities were taken from a large-scale professional photography database. • Synthetic identities created by a state-of-the-art open source generative model. • Same perceptual mechanisms govern natural and synthetic face discrimination. • Stimulus similarity and observers' ability impact natural and synthetic faces comparably. • Our findings highlight societal risks and research benefits of deepfakes.