[New Model] Chen2024Data

[CNOCycle]

Paper Information

• Paper Title: Data filtering for efficient adversarial training
• Paper URL: https://doi.org/10.1016/j.patcog.2024.110394
• Paper authors: Erh-Chung Chen, Che-Rung Lee

Leaderboard Claim(s)

Model 1

• Architecture: WRN-34-10
• Dataset: cifar10
• Threat Model: Linf
• eps: 8 / 255
• Clean accuracy: 86.54%
• Robust accuracy: 57.30%
• Additional data: false
• Evaluation method: AutoAttack
• Checkpoint and code: ckpt detail

Model 2

• Architecture: WRN-34-20
• Dataset: cifar10
• Threat Model: Linf
• eps: 8 / 255
• Clean accuracy: 86.10%
• Robust accuracy: 58.09%
• Additional data: false
• Evaluation method: AutoAttack
• Checkpoint and code: ckpt detail

Model 3

• Architecture: WRN-34-10
• Dataset: cifar100
• Threat Model: Linf
• eps: 8 / 255
• Clean accuracy: 64.32%
• Robust accuracy: 31.13%
• Additional data: false
• Evaluation method: AutoAttack
• Checkpoint and code: ckpt detail

Model 4

• Architecture: ResNet-50
• Dataset: imagenet
• Threat Model: Linf
• eps: 4 / 255
• Clean accuracy: 69.76%
• Robust accuracy: 41.92%
• Additional data: false
• Evaluation method: AutoAttack
• Checkpoint and code: ckpt detail

Model Zoo:

• [X] I want to add my models to the Model Zoo (check if true)
• [ ] I use an architecture that is not included among those here ( check if true).
• [X] I agree to release my model(s) under MIT license (check if true) OR
• [ ] I want my models to be released under a custom license, located here: (custom license URL here)

[fra31]

Hi,

thanks for the submission. I was adding the models (see https://github.com/RobustBench/robustbench/commit/31b7aa166b9d41eac9a42915e0fcfdeda383e8a7), and for the ImageNet one, it seems that the checkpoint corresponds to a WideResNet-50-2, is that right? Also, with the Res256Crop224 pre-processing I get 68.76% accuracy: are you maybe using another one?

[CNOCycle]

Hi.

Thank for your reminder. After checking the code for imagenet evaluation, the model architecture I used was the ResNet architecture provided by Robustbench but the width was set to width_per_group=64*2 manually. It should be WideResNet-50-2. Also, the pre-processing was Res256Crop224.

Thank you for clarifying that for me.

[fra31]

I still get 1% lower clean accuracy that reported. It might also have to do with different torchvision versions.

[CNOCycle]

I think that the subset I used for the evaluation might be different than the standard set provided by RobustBench. Besides, the evaluation was conducted almost two years ago, the version of AA or RobustBench I used seems to be older version. Variations (2%) in accuracy are acceptable.

[fra31]

Added the models with https://github.com/RobustBench/robustbench/pull/176, please let me know if there's anything to change.

[CNOCycle]

I greatly appreciate your assistance. Those commits appear satisfactory to me.