A fully differentiable pipeline for semantic adversarial examples targeting classification. Counterexamples are created with the help of Tzu-Mao Li's Redner software package.
This software package, written in PyTorch, allows for semantic attacks on an image constructed from a 3D model. More concretely, it takes an image classifier and a 3D model (in a .obj file format), renders the 3D model, and classifies it. You may attack the semantic features of the 3D model through one of 3 attributes: pose, vertex, or lighting. You can currently use one of 3 semantic attacks to do this -- our provided options are the Fast Gradient Sign Method, Projected Gradient Descent, and the Carlini-Wagner attack. More specific function details are concretely available in the code documentation in redner_adv.py.
Our experiments were done with the freely available ShapeNetCore dataset found here (you will have to make an account to download it).
A tutorial, using the enclosed demo motorcycle model, is available in scripts/tutorial.py, and the model checkpoint needed to execute the tutorial can be downloaded at the following link. Create a directory in your cloned repository folder called torch_models/ and save the checkpoint in it. The adversarial image is saved in the directory demo_out/ created by the script.