Research map / 3D vision / Gaussian splatting
AdLift: Lifting Adversarial Perturbations to Safeguard 3D Gaussian Splatting Assets Against Instruction-Driven Editing
Ziming Hong, Tianyu Huang, Runnan Chen, Shanshan Ye, Mingming Gong, Bo Han, Tongliang Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| initialize AdLift from 3DGS fitted on 2D adversarial perturbations (AdLift*) instead of from the raw asset (ablation) | AdLift-VU · IP2P editing, novel views (4 scenes avg) | CLIPs: 0.8181 -> 0.8087 | Improved |
| use untargeted VAE loss instead of targeted VAE loss (ablation) | AdLift* · IP2P editing, novel views (4 scenes avg) | CLIPs: 0.8296 -> 0.8087 CLIPs; 0.1758 -> 0.1802 CLIPd | Mixed |
| use L-PGD lifted perturbations rather than directly fitting 3DGS on per-view 2D protected images (ablation) | AdLift-VU · IP2P editing, training views (4 scenes avg) | CLIPs: 0.8368 -> 0.8139 | Improved |
| use L-PGD lifted perturbations rather than directly fitting 3DGS on per-view 2D protected images (ablation) | AdLift-VU · IP2P editing, novel views (4 scenes avg) | CLIPs: 0.8488 -> 0.8181 | Improved |
| use strict rendering-space bound (AdLift) instead of soft-constraint GaussianMarker at similar PSNR | Face scene protection · Face, novel views, matched PSNR | CLIPs: 0.8578 -> 0.7786 | Improved |
| perturb all Gaussian features (GuardSplat-Full) instead of SH features only (ablation) | GuardSplat · Face, novel views, matched PSNR | CLIPs: 0.8666 -> 0.8535 CLIPs at lower PSNR 25.7744 -> 23.7877 | No clear effect |
| apply AdLift*-VU protection trained on IP2P to unseen MagicBrush editing | AdLift*-VU · Face, MagicBrush, novel views | CLIPs: 0.8635 -> 0.7659 | Improved |
| apply AdLift*-VT protection to unseen SDEdit editing | AdLift*-VT · Face, SDEdit, novel views | CLIPd: 0.1708 -> 0.1751 | No clear effect |
| apply AdLift*-VU protection to unseen DGE 3D editing | AdLift*-VU · Face, DGE, novel views | CLIPd: 0.1218 -> 0.1200 | No clear effect |
| apply AdLift*-VU protection to unseen Instruct-GS2GS editing | AdLift*-VU · Face, Instruct-GS2GS, novel views | CLIPd: 0.1108 -> 0.0857 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GaussianMarker — GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting (NeurIPS 2024) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting (CVPR 2024 · similarity 0.88)
- EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D Gaussian Splatting (CVPR 2025 · similarity 0.87)
- DiffusionGuard: A Robust Defense Against Malicious Diffusion-based Image Editing (ICLR 2025 · similarity 0.87)
- BlurGuard: A Simple Approach for Robustifying Image Protection Against AI-Powered Editing (NeurIPS 2025 · similarity 0.86)
- SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography (ICLR 2025 · similarity 0.86)
- 3DGS-Drag: Dragging Gaussians for Intuitive Point-Based 3D Editing (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
