This directory contains all the code used in our experiments. It is organized into three main subdirectories:
- 3DTAB: This folder contains the code for evaluating 3D point cloud data.
- AT_old: This folder includes code for evaluating the results of adversarial training defenses from Ensemble Adversarial Training: Attacks and Defenses on ImageNet datasets.
- TransferAttack: This folder contains code for evaluating transferability in images.
Each folder requires a specific environment to run the corresponding code. You can create the required environment by running the install.sh script in each folder. The default environment name is the same as the folder name.
The experiments described in the paper are organized into scripts named run_exp_*.sh. The specific mapping of these scripts to the experiments can be found in the corresponding README.md file within each folder.
The model weights and dataset required for running the code can be found at bigfile. For detailed instructions on where to place the weights, refer to the README.md file in each respective folder.
For convenience, we provide all the adversarial examples generated during the experiments in TransferAttack, which can be download from here.
Our code is inspired by many excellent works. We sincerely thank the contributors of the following projects:
-
The benchmark setup and attack evaluation process are based on TransferAttack.
-
The implementation of iterative attack methods and registry design is based on ARES.
-
The attack pipeline for C&W-based methods draws from GeoA3 and 3d-adv-pc.
-
The evaluation code for models trained with ensemble adversarial training is adapted from admix.
-
The implementations of models, attack methods, and defense methods in 3DTAB reference the following repositories:
-
Models: PointNet, PointNet++, PointConv, PointCNN, CurveNet, Point-Transformers, PointMLP, Repsurf, PointCat, Pointcept, vnn
-
Attacks: PF-Attack, SI-Adv, HiT-ADV, ai-pointnet-attack, advpc
-
Defenses: IF-Defense, SI-Adv (provides a PyTorch implementation of the DUP-Net defense)
-
If you have any questions about the paper or the code, please contact zhazineedamail@163.com.