AdLift: Lifting 2D Adversarial Perturbations into 3D Gaussian Splatting Asset Protection Against Instruction-Driven Editing

AdLift proposes an editing protection scheme for 3D Gaussian Splatting (which the authors call the first), lifting strictly constrained 2D adversarial perturbations into protection for 3D Gaussian representations, preventing instruction-driven editing from taking effect across arbitrary viewpoints and dimensions. The method uses a modified Lifted PGD to alternately perform gradient clipping and image-to-Gaussian fitting on training views, writing the perturbation back into Gaussian parameters. Qualitative and quantitative results show that it defends against current state-of-the-art instruction-driven 2D image and 3DGS editing.