CoderXDY / AdvAttack

Zeroth-order non-convex optimization for Black-Box Adversarial Attacks against Deep Neural Networks

AdvAttack

Zeroth Order Optimization Based Black-box Attacks to Deep Neural Networks

This repository contains the code to implements various zero order algorithms used to perform adversarial attacks against black box Deep Neural Networks.
The full description of the problem, analysis and conclusion can be found in Report.pdf.

Requirements

• Python 3
• sys
• os
• time
• argparse
• warnings
• json
• datatime
• tqdm
• numpy
• matplotlib
• PIL
• sklearn
• torch
• torchvision

Code Structure

• papers
• models
• results
  • [optimizer]
  • Summary Results.ipynb
• scripts
  • test.py
  • evaluation.py
• src
  • dataset.py
  • models.py
  • train.py
  • loss.py
  • ZOOptim.py
  • zeroOptim.py
  • FWOptim.py
• Attack[dataset]_[optimizer].py
• TransferLearning.ipynb
• Report.pdf

Description

Papers contains the 4 papers used as references in this project.

Models contains the three models used and their weights: VGG16, InceptionV3 and an ad-hoc network.

Results contains one sub-folder for each Optimizer, each one containing the csv file related to the evaluation of the specified optimizer. It also contains a jupyter notebook used to generate the plots presented in the report.

Scripts contains two python files: test.py is used to perform an attack to a single image with the specified Optimizer, evaluation.py does the same on a batch of images.

Src contains:

• dataset.py is used to retrieve MNIST and CIFAR, to display some informations and eventually to rescale them to be compatible with InceptionV3.
• models.py is used to retrieve or create the three networks used (VGG16, InceptionV3 and an Ad-Hoc Network), to train and evaluate them.
• train.py is used to train the networks.
• loss.py implements the two loss functions used (Mean Squared Erorr and ZOO Loss).
• [XXX]Optim.py implement the optimizers described in the papers.

Attack[dataset]_[optimizer].py are used to perform a specific attack on a specific dataset and to visualize the generated adversarial image and the loss function.

Examples

Here we give an example of how to run an evaluation of Zero Stochastic Conditional Gradient with Inexact Updates against VGG16 fine-tuned on Cifar10. Only a subset of arguments will be given, full explanation of the arguments can be found inside the scripts.
If the data is not already present in the folder data the scripts will automatically create the folder data and download the data in there.

  cd scripts
  evaluation.py --optimizer "inexact" -- data "cifar10" --maximise 0 --epochs 100 --n_gradient 4000 --batch_size 1000 --mu 0.0025 --gamma 1       

N.B
1- All the scripts and jupyter notebook should be able to run without a problem in GPU thanks to cuda interface. If you don't have cuda or you are not able to have it (you don't have an NVIDIA GPU), we strongly suggest you to run everything on Google Colab.

2- The models uploaded are only VGG16 and MNISTNet while the InceptionV3 fine-tuned on Cifar10 and MNIST data are not present. To fine -tune your InceptionV3 model you can see TransferLearning.ipynb or you can contact us at [e-mail] and we will send you the model through Google Drive.