Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
DEGauss: Defending Against Malicious 3D Editing for Gaussian Splatting
Authors: Lingzhuang Meng, Mingwen Shao, Yuanjian Qiao, Xiang Lv
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our DEGauss not only performs excellent defense in different scenes, but also exhibits strong generalization across various state-of-the-art 3D editing pipelines. Moreover, Table 1 quantitatively evaluates our DEGauss against existing methods. It can be seen that our DEGauss achieves the highest PSNR, indicating that the introduced perturbations are more imperceptible. Regarding CLIP-based scores, our DEGauss attains the lowest similarity between the edited and original samples, demonstrating its effectiveness in disrupting malicious editing trajectories. |
| Researcher Affiliation | Academia | Lingzhuang Meng1, Mingwen Shao2 , Yuanjian Qiao3, Xiang Lv1 1 Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software, Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China), China 2 Artificial Intelligence Research Institute, Shenzhen University of Advanced Technology, China 3 College of Computer Science (College of Software), Inner Mongolia University, China EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology in Section 3 and mentions 'More settings and algorithm are provided in the Supplementary.' but does not include any pseudocode or algorithm blocks within the main body of the paper. |
| Open Source Code | No | We will open access to the code after the paper is accepted. |
| Open Datasets | Yes | Datasets and Editing Models. We verified the effectiveness on common 3D editing dataset [11, 32, 1], including face , girl , person , bear , bicycle , and garden scenes, with varying viewpoints and data scales. |
| Dataset Splits | No | The paper mentions using 'common 3D editing dataset [11, 32, 1], including face , girl , person , bear , bicycle , and garden scenes, with varying viewpoints and data scales' and that 'The total number of training steps is set to 2,000.', but it does not specify any explicit training, validation, or test dataset splits, percentages, or sample counts. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA RTX 4090 GPU. |
| Software Dependencies | No | Existing mainstream of 3D editing with 3DGS employ pretrained 2D diffusion models (e.g., Instruct Pix2Pix [2], Control Net [37]) as powerful generative priors to supervise the editing results. [...] However, the paper does not specify version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | We set the number of sampled views N = 6, the weighting factor τ = 1e-6, and γ = 1.0. To balance loss terms, we set hyperparameters λFD = λGD = 1e-5. The total number of training steps is set to 2,000. More settings and algorithm are provided in the Supplementary. |