ComaTransfer

Spectral Adversarial Geometric Attack on CoMA

This submodule is used to apply SAGA's adversarial shapes on CoMA. The code is inspired by CoMA's public implementation. Please follow the license rights of the authors if you use the code.

Installation

This code is tested on Python 3.6 and Pytorch version 1.3. Requirements can be installed by running:

conda create --name saga_transfer python=3.6 --yes
conda activate saga_transfer
bash install_transfer_requirements.sh

Please also install mesh processing libraries from MPI-IS/mesh. The steps are:

git clone https://github.com/MPI-IS/mesh.git
sudo apt-get install libboost-dev
cd mesh
pip install --upgrade pip
make all
cd ../

Prepare the Data

• Please copy the same training dataset as used for the victim autoencoder:

mkdir raw_data
cp ../coma/data/raw/coma_FEM.mat raw_data/
cp ../coma/data/raw/mesh_faces.npy raw_data/

• To test the trained model on SAGA's adversarial examples, please copy the desired results file from coma/results/ and place it under ComaTransfer/attack_data/. Then, rename it to attack_data_in.pickle .

Train the Model

To train the model please use:

python main.py

Test the Model

To evaluate the model please set the eval_flag to true and provide a checkpoint file. For example:

python main.py --eval_flag true --checkpoint_file_name checkpoint_sgd_0.008_199.pt 

Transfer SAGA

To evaluate and save the outputs of SAGA's adversarial shapes, please use:

python main.py --eval_attack_flag true --checkpoint_file_name checkpoint_sgd_0.008_199.pt 

The output is a dictionary, saved as attack_data_out.pickle under ComaTransfer/attack_data/.