Alternating Objectives Generates Stronger PGD-Based Adversarial Attacks

This is the implementation of our paper called "Alternating Objectives Generates Stronger PGD-Based Adversarial Attacks" [arXiv link].

---

Instructions

• First, install a virtual environment with the required dependencies and create a directory to save the datasets. Then, execute the preparation script, which downloads all the robust models of our study from the RobustBench ModelZoo.

python3 -m venv env 
source env/bin/activate 
pip install -r requirements.txt
mkdir Datasets # create a directory to save the datasets 
python3 code/prep_models_datasets.py

• Experiments can be run in three datasets: CIFAR-10, CIFAR-100 and ImageNet. Datasets are loaded inside the code/eval.py script. By default, RobustBench downloads CIFAR-10 and CIFAR-100 if the datasets do not exist in the prespecified path. However, the ImageNet val set has to be downloaded manually and then post-processed to have the expected directory structure. In order to fix this, follow the instructions from the relevant section at the RobustBench repository.
  Important: The datasets will be loaded to the path_to_dir argument of code/eval.py, which equals to the datasets_folder argument of run_eval.py script (by default, it will be in a directory called Datasets). If you've already obtained the ImageNet val directory, just create a soft link to the Datasets dir:

ln -s path/to/imagenet_val ./Datasets/val 

• The run_eval.py script can be used to evaluate the linf-bounded robustness of any model through PGD. For example, the following line evaluates the Engstrom2019Robustness model from RobustBench ModelZoo. In this case, PGD optimizes the CW loss with a step size of $\alpha = 0.25 * \epsilon$, remaining fixed during the process. The iteration budget is T = 100.

python3 run_eval.py --model_name Engstrom2019Robustness --loss CW --alpha_eps_ratio 0.25 --step_schedule None --iterations 100 --dataset CIFAR10

Results will be printed to the standard output and saved as .json files in the code/results directory.

Toy Example

In the code/toy_example directory, one can find a notebook reproducing the paper's toy example.