Isometric 3D Adversarial Examples in the Physical World

Authors: yibo miao, Yinpeng Dong, Jun Zhu, Xiao-Shan Gao

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments to evaluate the performance of our method on attacking typical point cloud recognition models [48, 49, 68]. Results demonstrate that, in comparison with the alternative state-of-the-art attack methods [63, 69, 88], ϵ-ISO attack achieves higher success rates, while making the generated adversarial examples more natural and robust under physical transformations. A physical-world experiment is conducted by 3D-printing the adversarial meshes and re-scanning the objects for evaluation, which also validates the effectiveness of our method.
Researcher Affiliation Collaboration Yibo Miao1,3 , Yinpeng Dong2,3 , Jun Zhu2,3,4,5, Xiao-Shan Gao1 1 KLMM, UCAS, Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China 2 Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua-Bosch Joint ML Center, THBI Lab, BNRist Center, Tsinghua University, Beijing 100084, China 3 Real AI 4 Peng Cheng Laboratory 5 Pazhou Laboratory (Huangpu), Guangzhou, China
Pseudocode Yes As shown in the overall algorithm (in Appendix B), we update the Bayesian posterior distribution of the objective f using the observations obtained from the previous iterations in the gradient descent process.
Open Source Code Yes The code is included in the supplementary material.
Open Datasets Yes We use the Model Net40 [71] dataset in our experiments. This dataset contains 12,311 CAD models with 40 common object semantic categories in the real world. We use the official split [48, 49], where 9,843 examples are used for training and the remaining 2,468 examples are used for testing.
Dataset Splits No The paper specifies training and testing splits, but does not explicitly mention a separate validation dataset split.
Hardware Specification No The paper mentions physical hardware used for scanning and printing: 'The selected meshes are printed by the Stratsys J850 Prime 3D printer and scanned by the Ein Scan-SE 3D scanner.' It also mentions 'High Performance Computing Center, Tsinghua University' in the acknowledgements. However, it does not provide specific details about the computational hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions using 'Adam [31] to optimize the objective' but does not specify version numbers for any software libraries, programming languages, or other dependencies.
Experiment Setup Yes We use a fixed learning schedule of 250 iterations, where the learning rate and momentum are respectively set as 0.01 and 0.9. We assign the weighting parameters λ1 = 1.0, λ2 = 0.2 and λ3 = 0.8. The balancing parameter β is initialized as 1,500 and automatically adjusted by conducting 10 runs of binary search following [4].