[New Model] <Huang2022Revisiting>

[zhichao-lu]

Paper Information

• Paper Title: Revisiting Residual Networks for Adversarial Robustness: An Architectural Perspective
• Paper URL: https://arxiv.org/abs/2212.11005
• Paper authors: Shihua Huang, Zhichao Lu, Kalyanmoy Deb, Vishnu Naresh Boddeti

Leaderboard Claim(s)

Add here the claim for your model(s). Copy and paste the following subsection for the number of models you want to add.

Model 1

• Architecture: WRN-A4
• Threat Model: Linf
• eps: (N/A for common corruptions)
• Clean accuracy: 91.58
• Robust accuracy: 65.79
• Additional data: true
• Evaluation method: AutoAttack
• Checkpoint and code: codes (https://github.com/zhichao-lu/robust-residual-network), checkpoints (https://pan.baidu.com/s/1t2ZFXf_3DoKX_W7j4nzxuw?pwd=78cb)

Model Zoo:

• [x] I want to add my models to the Model Zoo (check if true)
• [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 (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,

sorry for the late reply, we'll add the models as soon as possible!

[fra31]

I'm having some trouble trying to download the checkpoint. Also, there seem to be several checkpoints in the linked folder. Could you please provide a direct link to the download?

[zhichao-lu]

Sorry for the inconvenience, could you please try this link (https://drive.google.com/file/d/1Z7R_dZw-Ds0ofc7eqnHagIKNN2dnlJtU/view?usp=sharing)?

[fra31]

This works fine, thanks! However, I'm getting only 81.50% clean accuracy. This is the model I'm using following the setup here:

stride_config = [1, 2, 2]
activations = ('ReLU', 'ReLU', 'ReLU')
normalizations = ('BatchNorm', 'BatchNorm', 'BatchNorm')
depth, width_mult = [27, 28, 13], [10, 14, 6]
block_types = ['basic_block', 'basic_block', 'basic_block']
scales, base_width, cardinality, se_reduction = None, None, None, None
channels = [16, 16 * width_mult[0], 32 * width_mult[1], 64 * width_mult[2]]
model = PreActResNet(
    num_classes=10,
    channel_configs=channels,
    depth_configs=depth,
    stride_config=stride_config,
    stem_stride=1,
    block_types=block_types,
    activations=activations,
    normalizations=normalizations,
    use_init=True,
    cardinality=cardinality,
    base_width=base_width,
    scales=scales,
    se_reduction=se_reduction,
    pre_process=True,
)

Is there any different configuration to use (the checkpoint can be successfully loaded)?

[zhichao-lu]

Could you please try this configuration file (https://drive.google.com/file/d/11cIXeMln7EQQaWkRPkyiN_bTKf9ku6pv/view?usp=sharing)?

Or simply change the activations to activations = ('SiLU', 'SiLU', 'SiLU')

[fra31]

Changing the activations yields 91.42%, which is still a bit lower that what you report. I think the other parameters match your configuration, but please let me know if there's something else to change (here the current implementation).

[zhichao-lu]

It turns out that I shared the wrong checkpoint, sorry about that.

Here is the correct checkpoint (https://drive.google.com/file/d/1PSe1LOvX6r3_X6Vvc6oNYh-q4C10JfpJ/view?usp=sharing), and here is the output log we got (https://drive.google.com/file/d/14_O_2Dmt9PFoQlH2aEfNMYXz7C5UqX5G/view?usp=sharing)

Please use silu with this checkpoint.

Sorry again for all the inconvenience.

[fra31]

No problem, the new checkpoint works fine, thanks!

[fra31]

Added the model with https://github.com/RobustBench/robustbench/pull/125, please let me know if there's something to adjust.

[fra31]

The leaderboard is updated as well https://github.com/RobustBench/robustbench.github.io/commit/e5f57c9d6a67cc5adcb803065cad312b564da78f.