Explainability-Aware One Point Attack for Point Cloud Neural Networks

Pytorch implementation for Explainability-Aware One Point Attack for Point Cloud Neural Networks. Point cloud neural networks are based on this repo. Please follow the instructions to train networks before attacking them.

Environments

Python >= 3.6 Pytorch >= 1.6.0

Usage

Before running the code, move the test data file list modelnet40_test_adv.txt to ./data/modelnet40_normal_resampled/ or generate user-defined number of test files using sample_adv_test_data.py (also should be placed in the ./data/modelnet40_normal_resampled/ path):

python sample_adv_test_data.py

Create visualization path:

mkdir visu
cd visu
mkdir output

Test and visualize one instance randomly picked up from dataset with OPA and CTA respectively:

python Test_single_ins_OPA.py
python Test_single_ins_CTA.py

Quantitatively evaluate the attack performance:

python Eval_OPA.py
python Eval_CTA.py