Double Bubble, Toil and Trouble: Enhancing Certified Robustness through Transitivity

This code contains the numerical materials required to replicate the 2022 Neurips paper (accepted) Double Bubble, Toil and Trouble: Enhancing Certified Robustness through Transitivity.

License

Published under an MIT license

To support Figure 5 of the published work, a subset of the published code is drawn from the MACER libaries of https://github.com/RuntianZ/macer - all pieces of code based upon MACER have this attribution indicated in the preamble of the code.

Repository

Outline

• src/ contains all code required to run the experiments
• data/ contains the trained models. Due to their size, trained models have been excluded
• Example bash scripts to run evaluation and analysis are included in *.sh files

Dependencies

• Python3, PyTorch with CUDA, Torchvision, PIL, Numpy, Matplotlib, SciPy, Statsmodels