Evaluating GAIRAT robustness using Logit Scaling Attack. We evaluate the adversarial robustness of a very recent method called "Geometry-aware Instance-reweighted Adversarial Training" using Logit Scaling Attack.
The results of our experiments can be found here.
To test GAIRAT on CIFAR-10 we had to train their model and our pre-trained models can be found here.
Download our pre-trained model.
Then:
pip install tqdm torch==1.7.1+cu110 torchvision==0.8.2+cu110 -f https://download.pytorch.org/whl/torch_stable.html
to install the needed dependencies. We tested using PyTorch 1.7.1 and CUDA 11.0.
python eval_pgd.py --model_path <model_path> --output_suffix=<result_path> --num_restarts 1 --num_steps 20 --alpha <alpha>
to test the model using different values of alpha (the scaling factor).
<model_path>is the pre-trained model path (e.g. checkpoint.pth.tar).<result_path>is the path where to store evaluation results.<alpha>is the desired scaling value (e.g. 10.0).