Home

Awesome

Certified Patch Robustness via Smoothed Vision Transformers

This repository contains the code for replicating the results of our paper:

Certified Patch Robustness via Smoothed Vision Transformers </br> Hadi Salman*, Saachi Jain*, Eric Wong*, Aleksander Madry

Paper <br> Blog post Part I. <br> Blog post Part II. <br>

    @article{salman2021certified,
        title={Certified Patch Robustness via Smoothed Vision Transformers},
        author={Hadi Salman and Saachi Jain and Eric Wong and Aleksander Madry},
        booktitle={ArXiv preprint arXiv:2110.07719},
        year={2021}
    }
<p> <img src="main_fig.png" width="1000" > </p>

Getting started

Our code relies on the MadryLab public robustness library, which will be automatically installed when you follow the instructions below.

  1. Clone our repo: git clone https://github.mit.edu/hady/smoothed-vit

  2. Install dependencies:

    conda create -n smoothvit python=3.8
    conda activate smoothvit
    pip install -r requirements.txt
    

Full pipeline for building smoothed ViTs.

Now, we will walk you through the steps to create a smoothed ViT on the CIFAR-10 dataset. Similar steps can be followed for other datasets.

The entry point of our code is main.py (see the file for a full description of arguments).

First we will train the base classifier with ablations as data augmentation. Then we will apply derandomizd smoothing to build a smoothed version of the model which is certifiably robust.

Training the base classifier

The first step is to train the base classifier (here a ViT-Tiny) with ablations.

python src/main.py \
      --dataset cifar10 \
      --data /tmp \
      --arch deit_tiny_patch16_224 \
      --pytorch-pretrained \
      --out-dir OUTDIR \
      --exp-name demo \
      --epochs 30 \
      --lr 0.01 \
      --step-lr 10 \
      --batch-size 128 \
      --weight-decay 5e-4 \
      --adv-train 0 \
      --freeze-level -1 \
      --drop-tokens \
      --cifar-preprocess-type simple224 \
      --ablate-input \
      --ablation-type col \
      --ablation-size 4

Once training is done, the mode is saved in OUTDIR/demo/.

Certifying the smoothed classifier

Now we are ready to apply derandomized smoothing to obtain certificates for each datapoint against adversarial patches. To do so, simply run:

python src/main.py \
      --dataset cifar10 \
      --data /tmp \
      --arch deit_tiny_patch16_224 \
      --out-dir OUTDIR \
      --exp-name demo \
      --batch-size 128 \
      --adv-train 0 \
      --freeze-level -1 \
      --drop-tokens \
      --cifar-preprocess-type simple224 \
      --resume \
      --eval-only 1 \
      --certify \
      --certify-out-dir OUTDIR_CERT \
      --certify-mode col \
      --certify-ablation-size 4 \
      --certify-patch-size 5

This will calculate the standard and certified accuracies of the smoothed model. The results will be dumped into OUTDIR_CERT/demo/.

That's it! Now you can replicate all the results of our paper.

Download our ImageNet models

If you find our pretrained models useful, please consider citing our work.

Models trained with column ablations

ModelAblation Size = 19
ResNet-18LINK
ResNet-50LINK
WRN-101-2LINK
ViT-TLINK
ViT-SLINK
ViT-BLINK

Models trained with block ablations

ModelAblation Size = 75
ViT-TLINK
ViT-SLINK
ViT-BLINK

We have uploaded the most important models. If you need any other model (for the sweeps for example) please let us know and we are happy to provide!

Maintainers