Researchers discover critical security vulnerabilities in fast 3D reconstruction AI models used for real-time 3D visualization

AdvSplat introduces the first systematic study of adversarial attacks targeting feed-forward 3D Gaussian Splatting (3DGS) models

Feed-forward 3DGS models enable fast 3D reconstruction from minimal input views without scene-specific optimization, but this speed advantage comes with security risks

Researchers demonstrate white-box attacks revealing fundamental vulnerabilities, then develop query-efficient black-box algorithms that exploit pixel-space perturbations

Neural network backbones in these models increase susceptibility to adversarial manipulation despite their commercial deployment potential