This repository is the official implementation of A Unified Game-Theoretic Interpretation of Adversarial Robustness.

pip install -r requirements.txt

We used the pre-trained models on the ImageNet dataset released by https://github.com/microsoft/robust-models-transfer.

Here, we take the untargeted PGD attack on the standard ResNet-18 as an example.

1. Generate adversarial examples on the pre-trained model

   python gene_adv_imgs_untarget.py --root=<path_to_data> --dataset_name=imagenet --arch=resnet18 --std_model_path=<path_to_model>

2. Sample variable pairs and contexts for the computation of interactions

   python gene_pairs.py --arch=resnet18 --ratios 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

3. Compute the model output on masked samples

   python m_order_interaction_logit.py --root=<path_to_data> --dataset_name=imagenet --arch=resnet18 --std_model_path=<path_to_model> --ratios 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

4. Compute interactions based on the model output

   python compute_interactions.py --root=<path_to_data> --dataset_name=imagenet --arch=resnet18 --std_model_path=<path_to_model> --ratios 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

5. Plot the average interaction in normal samples and adversarial examples, just like Figure 3 in the paper.

   python plot.py --arch=resnet --ratios 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9