This repository contains the implementation of adversarial attack methods targeting MedShapeNet's 3D meshes. The code is designed to manipulate and perturb the mesh data in a way that degrades the performance of machine learning models trained on this dataset, specifically reducing their classification accuracy. The generated adversarial examples can be used for further research into the robustness and security of mesh-based neural networks, aiding in the development of more resilient models.

Created by Yuval Rosman and Ben Volovelsky.

This repository contains the implementation of MeshWalker and Random-Walks-for-Adversarial-Meshes.

Note for this README: replace <dataset> with medshapenet.

To get the raw datasets go to the relevant website, and put it under MeshWalker/datasets_raw/medshapenet.

• MedShapeNet.

To prepare the data, run python dataset_prepare.py medshapenet.

Or request the data directly from us.

Make sure to put the processed dataset in MeshWalker/datasets_processed/medshapenet. Processing will rearrange dataset in npz files, labels included, vertex neighbors added.

python train_val.py medshapenet

You will find the results at: MeshWalker\runs\???

Use tensorboard to show training results: tensorboard <trained-model-folder>

Note that "accuracy" tab is a fast accuracy calculated while training, it is not the final accuracy we get using averaging. To get the final accuracy results, please refer to the "full_accuracy" tab at tensorboard, or run evaluation scripts.