Adversarial attacks remain one of the main flaws of deep neural networks today. Besides adversarial training, some researchers have identified derivative regularisation (first and second order) to be another usefool tool for both understanding and tackling adversarial robustness. However, these methods do not quite reach the same level of effectiveness as adversarial training yet.

In this repository, we explore the role of mixed and higher order effects in regularization and how/if considering them can further narrow the gap to adversarial training. Furthermore, we explore improving the accuracy of derivative estimators to higher orders.

The base for our code is https://github.com/F-Salehi/CURE_robustness. However, we have changed quite a lot, so this repo has quite different dependencies.

To run an example execution of our code, run CURE_robustness_mod/main.py. You can set hyper parameters for e.g. regularizer constants, data set and level of estimation accuracy in CURE_robustness_mod/utils/config.py.

To reproduce our experiments, run CURE_robustness_mod/reproduce_experiments.py. You are advised to use a sufficiently strong GPU with at least 8GB of memory.

We recommend you to create a conda or virtual environment from our CURE_robustness_mod/requirements.txt with python version 3.7.1. The code might still run if you do it differently, but no guarantees. 🙅

If you still have any questions, feel free to reach out to us. Otherwise, enjoy our code! 🙌