Toward Availability Attacks in 3D Point Clouds
Authors: Yifan Zhu, Yibo Miao, Yinpeng Dong, Xiao-Shan Gao
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on typical point cloud datasets, 3D intracranial aneurysm medical dataset, and 3D face dataset verify the superiority and practicality of our approach. |
| Researcher Affiliation | Collaboration | 1Academy of Mathematics and Systems Science, Chinese Academy of Sciences 2University of Chinese Academy of Sciences 3Tsinghua University 4Real AI 5Kaiyuan International Mathematical Sciences Institute. Correspondence to: Xiao-Shan Gao <xgao@mmrc.iss.ac.cn>, Yibo Miao <miaoyibo@amss.ac.cn>. |
| Pseudocode | Yes | Algorithm 1 Feature Collision Error-Minimization Poisoning Attack (FC-EM) Input: A 3D point cloud training dataset D = {(xi, yi)}N i=1. Total epoch T. Batch size NB. Distance loss Ldis and regularization strength β. Feature collision loss Lfc and temperature t. Classifier parameters αθ and Tθ. Attack parameters αδ, Tδ and Ta. Output: Poisoned dataset Dδ = {(xi + δi, yi)}N i=1 Initialize: δi 0, i = 1, 2, , N for t = 1, , T do Sample a mini batch B = {(xbj, ybj)}NB j=1. θ θ αθ θE(xbj ,ybj ) B Lcls(xbj + δbj, ybj; θ) + β Ldis(xbj + δbj, xbj) for tθ = 1, , Tδ do Sample a mini batch B = {(xbj, ybj)}NB j=1. for ta = 1, , Ta do Compute class-wise feature collision loss Lfc. δbj δbj αδ δbj E(xbj ,ybj ) B Lfc(xbj +δbj, ybj; θ, t) + β Ldis(xbj + δbj, xbj) |
| Open Source Code | Yes | Code is available at https://github.com/hala64/fc-em. |
| Open Datasets | Yes | Dataset. In our experiments, we employ several datasets for evaluation, including the generated point clouds dataset, Model Net40 (Wu et al., 2015), the scanned point clouds dataset, Scan Object NN (Uy et al., 2019), the 3D intracranial aneurysm medical dataset, Intr A (Yang et al., 2020), and the BFM-2017 generated face dataset (Gerig et al., 2018). Details of these datasets are provided in Appendix C. |
| Dataset Splits | Yes | Model Net40 (Wu et al., 2015) is a collection of CAD models with 40 object categories, which is divided into 9843 training data and 2468 test data. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware (e.g., GPU models, CPU types, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using PyTorch but does not provide specific version numbers for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | Implementation details. We use Py Torch learning rate scheduler, Reduce LROn Plateau for both poisoning and evaluation process, with initial learning rate 10 3. Except for above baseline methods, we also evaluate the targeted version of AP and REG-AP, namely AP-T and REG-AP-T. The poisoning epoch for EM, REG-EM and FC-EM is set to 200, and for AP(-T), REG-AP(-T) is set to 100. The evaluation epoch is set to 200. The coefficient of regularization term β is set to 1.0 originally. Across all datasets, we uniformly sample 1,024 points for attacks and classification. For further details, including more ablation studies and experiments, please refer to Appendices C and D. |