[New Model] RodriguezMunoz2024Characterizing

[adriarm]

Paper Information

• Paper Title: Characterizing Model Robustness via Natural Input Gradients
• Paper URL: https://arxiv.org/pdf/2409.20139
• Paper authors: Adrián Rodríguez-Muñoz, Tongzhou Wang, Antonio Torralba

The focus of the paper is on efficient robust training using gradient regularisation, and showing it is surprisingly effective on ImageNet.

Leaderboard Claim(s)

Our model can be instantiated and benchmarked by:

1. Cloning our code and following environment install instructions.
2. Downloading our checkpoint
3. Setting data_dir, arch and ckpt_location in benchmark_model.py
4. Running benchmark_model.py

The models use the Swin architecture from the old timm==0.6.7 version. If it is preferred for me to write the pull request, please let me know what the best way to implement it is.

Model 1

• Architecture: Swin-B (from timm==0.6.7, same as Liu2023Comprehensive)
• Threat Model: Linf
• eps: 4 / 255
• Clean accuracy: 77.76
• Robust accuracy: 51.56
• Additional data: False
• Evaluation method: AutoAttack
• Checkpoint and code: Checkpoint and code.

Model 2

• Architecture: Swin-L (from timm==0.6.7, same as Liu2023Comprehensive)
• Threat Model: Linf
• eps: 4 / 255
• Clean accuracy: 79.36
• Robust accuracy: 53.82
• Additional data: False
• Evaluation method: AutoAttack
• Checkpoint and code: Checkpoint and code.

Model Zoo:

• I want to add my models to the Model Zoo (check if true)
• I use an architecture that is included among
  those here or in timm. If not, I added the link to the architecture implementation so that it can be added.
• I agree to release my model(s) under MIT license (check if true) OR under a custom license, located here: (put the custom license URL here if a custom license is needed. If no URL is specified, we assume that you are fine with MIT)

[fra31]

Hi,

thanks for the submission! I'll try to add the models in the next days.