Poison-splat: Computation Cost Attack on 3D Gaussian Splatting
Authors: Jiahao Lu, Yifan Zhang, Qiuhong Shen, Xinchao Wang, Shuicheng YAN
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | All experiments in this paper are carried out on three common 3D datasets: (1) Ne RFSynthetic2 (Mildenhall et al., 2021) is a synthetic 3D object dataset. (2) Mip-Ne RF3603 dataset (Barron et al., 2022) is a 3D scene dataset. (3) Tanks-and Temples4 dataset (Knapitsch et al., 2017). ... We use the number of 3D Gaussians, GPU memory occupancy and the training time cost as metrics to evaluate the computational cost of 3DGS. ... We report the comparisons of the number of Gaussians, peak GPU memory, training time and rendering speed between clean and poisoned data across three different datasets in Tables 1 to show the effectiveness of our attack. |
| Researcher Affiliation | Collaboration | Jiahao Lu1 Yifan Zhang2 Qiuhong Shen1 Xinchao Wang1 Shuicheng Yan2,1 1 National University of Singapore 2 Skywork AI |
| Pseudocode | Yes | Algorithm 1 Poison-splat Input: Clean dataset: D = {Vk, Pk}; Perturbation range: ϵ; Perturbation step size: η; The iteration number of inner optimization: T; The iteration number of outer optimization: T. Output: Poisoned dataset: Dp = { Vk, Pk} |
| Open Source Code | Yes | 1Our code is available at https://github.com/jiahaolu97/poison-splat. |
| Open Datasets | Yes | All experiments in this paper are carried out on three common 3D datasets: (1) Ne RFSynthetic2 (Mildenhall et al., 2021)...2Dataset publicly accessible at https://github.com/bmild/nerf. (2) Mip-Ne RF3603 dataset (Barron et al., 2022)...3Dataset publicly accessible at https://jonbarron.info/mipnerf360/ (3) Tanks-and Temples4 dataset (Knapitsch et al., 2017)...4Dataset publicly accessible at https://www.tanksandtemples.org/download/ |
| Dataset Splits | No | The paper mentions using common 3D datasets for experiments but does not explicitly provide details about training, validation, or test splits (e.g., percentages, absolute counts, or references to standard splits) for these datasets in the provided text. |
| Hardware Specification | Yes | All experiments reported in the paper were carried out on a single NVIDIA A800-SXM4-80G GPU. |
| Software Dependencies | No | The paper mentions using the "official implementation of 3D Gaussian Splatting" and "Scaffold-GS" but does not provide specific version numbers for these or any underlying software libraries (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | In our experiments, we use the official implementation of 3D Gaussian Splatting5 (Kerbl et al., 2023) to implement the victim behavior. Following their original implementation, we use the recommended default hyper-parameters, which were proved effective across a broad spectrum of scenes from various datasets. ... Algorithm 1 Poison-splat Input: Clean dataset: D = {Vk, Pk}; Perturbation range: ϵ; Perturbation step size: η; The iteration number of inner optimization: T; The iteration number of outer optimization: T. ... for unconstrained attacks on MIP-Ne RF360, we set T = 6000 and T = 25. |