Evaluating Privacy Risks of Facial Recognition on Photorealistic Avatars in Social Virtual Reality

Photorealistic avatars are increasingly adopted in Social Virtual Reality to enhance presence, yet their close resemblance to users may introduce biometric privacy risks through unwanted face recognition and re-identification. We assess whether state-of-the-art face recognition models can identify avatars derived from real individuals and whether offline visual obfuscation techniques - blurring, mesh deformation, and occlusion - can mitigate this risk. Under the evaluated dataset and threat model, avatars are consistently identifiable without obfuscation, whereas high-intensity facial occlusion substantially reduces identifiability.