Denoised Smoothing: A Provable Defense for Pretrained Classifiers

This repository contains the code and models necessary to replicate the results of our recent paper:

Denoised Smoothing: A Provable Defense for Pretrained Classifiers
Hadi Salman, Mingjie Sun, Greg Yang, Ashish Kapoor, J. Zico Kolter
NeurIPS 2020
Paper: https://arxiv.org/abs/2003.01908
Blog post: https://www.microsoft.com/en-us/research/blog/denoised-smoothing-provably-defending-pretrained-classifiers-against-adversarial-examples/

Our paper presents a method for provably defending any pretrained image classifier against Lp adversarial attacks.

Overview of the Repository

Our code is based on the open source codes of Cohen et al (2019) and Salman et al. (2019). The major contents of our repo are as follows:

• vision_api/ contains the code for our experiments on online Vision APIs. Check out the tutorial!

  # To robustify an ONLINE api (api_name can be "azure", "google", "aws", "clarifai")
  majority_class, _, _ = RobustAPI(api_name, denoiser=denoiser, online=True).predict(img, ...)
  majority_class, radius, logs = RobustAPI(api_name, denoiser=denoiser, online=True).certify(img, ...)
  
  # To use the OFFLINE version (i.e. read from previous query logs, no denoiser needed)
  majority_class, _ = RobustAPI(api_name, online=False).predict(logs, ...)
  majority_class, radius = RobustAPI(api_name, online=False).certify(logs, ...)
  

• code/ contains the code for our experiments on CIFAR-10 and ImageNet.

• analysis/ contains the plots and tables that are shown in our paper. Keep reading to see how you can replicate these easily!

Let us dive into the files in code/:

1. train_classifier.py: a generic script for training ImageNet/Cifar-10 classifiers, with Gaussian agumentation option, achieving SOTA.
2. train_denoiser.py: the main code of our paper which is used to train the different denoisers used in our paper.
3. train_denoiser_multi_classifier.py: a variant of train_denoiser.py that allows training denoisers using multiple surrogate models.
4. test_denoiser.py: a script to test the performance of the denoiser on reconstruction task, and on image classification under Gaussian noise when a pretrained classifier is attached to the denoiser.
5. visualize.py: a script for visualizing noisy images and denoised images.
6. certify.py: Given a pretrained smoothed classifier, returns a certified L2-radius for each data point in a given dataset using the algorithm of Cohen et al (2019).
7. architectures.py: an entry point for specifying which model architecture to use per classifiers and denoisers.

Getting Started

1. git clone https://github.com/microsoft/denoised-smoothing.git

2. Install dependencies:

   conda create -n denoised-smoothing python=3.6
   conda activate denoised-smoothing
   conda install numpy matplotlib pandas seaborn scipy==1.1.0
   conda install pytorch torchvision cudatoolkit=10.0 -c pytorch # for Linux
   pip install google-cloud-vision boto3 clarifai
   

3. Download and extract our certification logs from here. You can instead simply run the following from within the root directory of this repository

   wget -O data.tar.gz https://www.dropbox.com/s/fjmncwhsnfgkmzk/data.tar.gz?dl=0 && tar -xzvf data.tar.gz
   

4. Download our trained models (denoisers and classifiers) from here. Then move the downloaded pretrained_models.tar.gz into the root directory of this repository. Run tar -xzvf pretrained_models.tar.gz to extract the models.

5. If you want to run ImageNet experiments, obtain a copy of ImageNet and preprocess the val directory to look like the train directory by running this script. Finally, set the environment variable IMAGENET_DIR to the directory where ImageNet is located.

6. Let us try to certify the robustness of a CIFAR-10 pretrained model with an attached MSE-trained DnCNN denoiser.

   pretrained_classifier="pretrained_models/cifar10_classifiers/ResNet110_90epochs/noise_0.00/checkpoint.pth.tar"
   denoiser="pretrained_models/trained_denoisers/cifar10/mse_obj/dncnn/epochs_90/noise_0.25/checkpoint.pth.tar"
   output="certification_output/sigma_0.25"
   python code/certify.py --dataset cifar10 --base_classifier $pretrained_classifier --sigma 0.25 --outfile $output --skip 20 --denoiser $denoiser
   

Check the results in certification_output/sigma_0.25. You should get similar to data/certify/cifar10/mse_obj/MODEL_resnet110_90epochs_DENOISER_cifar10_dncnn_epochs_90/noise_0.25/test_N10000/sigma_0.25

Are they similar? Perfect! You can keep going.

Example

Let's now convert a pretrained non-robust CIFAR-10 classifier to a provably robust one!

In what follows, we will show you how you can train a denoiser on CIFAR-10 using the MSE objective, attach it to a pretrained classifier, then certify the robustness of the resultant robust classifier. This is the pretrained model we consider

pretrained_classifier="pretrained_models/cifar10_classifiers/ResNet110_90epochs/noise_0.00/checkpoint.pth.tar"

Training the denoiser

• To train a denoiser with MSE loss to denoise Gaussian noise of stddev of 0.25, run the following

python code/train_denoiser.py --dataset cifar10 --arch cifar_dncnn --outdir denoiser_output_dir --noise 0.25

Lazy to train? No worries, we have trained one for you! Just run the following in the command-line, and continue with the example

denoiser_output_dir=pretrained_models/trained_denoisers/cifar10/mse_obj/dncnn/epochs_90/noise_0.25 

Test the denoiser

Let's check how good the trained denoiser is,

python code/test_denoiser.py --dataset cifar10 --denoiser $denoiser_output_dir/checkpoint.pth.tar --clf $pretrained_classifier --noise 0.25

Certification

• Certify the trained model on CIFAR-10 test set using σ=0.25

python code/certify.py --dataset cifar10 --base_classifier $pretrained_classifier --sigma 0.25 
--outfile certification_output/sigma_0.25 --skip 20 --denoiser $denoiser_output_dir/checkpoint.pth.tar

will load the $denoiser and attach it to the pretrained classifier $pretrained_classifier, smooth it using a noise level σ=0.25, and certify 500 samples of the cifar10 test set.

If you check the results in certification_output/sigma_0.25, you should again get similar to data/certify/cifar10/mse_obj/MODEL_resnet110_90epochs_DENOISER_cifar10_dncnn_epochs_90/noise_0.25/test_N10000/sigma_0.25.

So what? What has just happened? In fact, you have just converted a pretrained CIFAR-10 model into a provably robust one (for each image of the CIFAR-10 test set, you have a certified L2 radius within which the prediction is constant!)

To see this more clearly, let's try to certify the pretrained classifier without using a denoiser and compare the certification results.

python code/certify.py --dataset cifar10 --base_classifier $pretrained_classifier --sigma 0.25 
--outfile certification_output/sigma_0.25_no_denoiser --skip 20

The outputcertification_output/sigma_0.25_no_denoiser should be something like
data/certify/cifar10/no_denoiser/MODEL_resnet110_90epochs/noise_0.00/test_N10000/sigma_0.25.

Now, run python code/generate_github_result.py (you might need to change the paths to the certification results in this script) to generate the below certification curves from the above certification results, you will get

Note how adding a denoiser substantially improves the certified accuracy of the pretraing classifier!

Replicate our Tables and Figures

We provide code to generate all the tables and results of our paper. Simply run

python code/analyze.py

This code reads from the data/ folder (which should appear if you followed the Getting started section correctly) i.e. the logs that were generated when we certifiied our trained models, and automatically generates the tables and figures that we present in the paper.

Below are example plots from our paper which you will be able to replicate by running the above code.

Download our Pretrained Models

You can download our trained models here. These contain all our trained denoisers and pretrained classfiers that we use in our paper.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repositories using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.