AI-Powered Facial Mask Removal Is Not Suitable For Biometric Identification
For law enforcement and social media platforms, it demonstrates that AI-unmasked images are unsuitable for biometric identification, addressing a critical safety concern.
The paper evaluates commercial AI-powered facial unmasking tools and finds that generated faces cannot be reliably matched to true identities, with matching accuracy below 10% in controlled tests.
Recently, crowd-sourced online criminal investigations have used generative-AI to enhance low-quality visual evidence. In one high-profile case, social-media users circulated an "AI-unmasked" image of a federal agent involved in a fatal shooting, fueling a wide-spread misidentification. In response to this and similar incidents, we conducted a large-scale analysis evaluating the efficacy and risks of commercial AI-powered facial unmasking, specifically assessing whether the resulting faces can be reliably matched to true identities.