[New Model] <bartoldson2024adversarial>

[bbartoldson]

Paper Information

• Paper Title: Adversarial Robustness Limits via Scaling-Law and Human-Alignment Studies
• Paper URL: https://arxiv.org/abs/2404.09349
• Paper authors: Brian Bartoldson, James Diffenderfer, Konstantinos Parasyris, Bhavya Kailkhura

Leaderboard Claim(s)

Model 1 is the CIFAR-10 SOTA for Linf. Model 2 has only 79M parameters.

Both architectures can be instantiated via the DMWideResNet class shown in the lambda function here, replacing width and depth with those specified in the architecture names below (we use SiLU and the mean/std shown in the linked example). Our code is not yet public, but it will be posted at the link provided below soon.

Model 1

• Architecture: WRN-94-16
• Threat Model: Linf
• eps: 8/255
• Clean accuracy: 93.68%
• Robust accuracy: 73.71%
• Additional data: false
• Evaluation method: AutoAttack
• Checkpoint and code: checkpoint and code

Model 2

• Architecture: WRN-82-8
• Threat Model: Linf
• eps: 8/255
• Clean accuracy: 93.11%
• Robust accuracy: 71.59%
• Additional data: false
• Evaluation method: AutoAttack
• Checkpoint and code: checkpoint and code

Model Zoo:

• [x] I want to add my models to the Model Zoo
• [x] 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.
• [x] I agree to release my model(s) under MIT license

[fra31]

Hi,

thanks for the submission. I've added the models with https://github.com/RobustBench/robustbench/pull/185. Please let me know if there's anything to modify.

[bbartoldson]

Hi,

This looks great, and I have a couple of minor suggestions. The following footnotes currently reflect the number of synthetic images trained on, but I think they should reflect the number of unique synthetic images trained on.

Footnote for WRN-82-8

• current: "It uses additional 349M synthetic images in training."
• suggestion: "It uses additional 300M synthetic images in training."

Footnote for WRN-94-16

• current: "It uses additional 500M synthetic images in training.",
• suggestion: "It uses additional 300M synthetic images in training."

Thanks!

Brian

[fra31]

Updated the footnotes (https://github.com/RobustBench/robustbench/pull/185/commits/c809251f1ac6192cb6a467204055fa3c73ef084f).