Pytorch implementation NAG - Network for Adversary Generation

Official Project Page: Link

Authors: Konda Reddy Mopuri*, Utkarsh Ojha*, Utsav Garg, R. Venkatesh Babu.

This work is an attempt to explore the manifold of perturbations that can cause CNN based classifiers to behave absurdly. At present, this repository provides the facility to train the generator that can produce perturbations to fool VGG F, VGG 16, VGG 19, GoogleNet, CaffeNet, ResNet 50, ResNet 152. The generator architecture has been modified from here.

Proposed Approach

The core idea is to model the distribution of adversarial perturbations using a Generative approach, where in the discriminator used is a pretrained model, In this approach the Only genertors is getting updated. To Quantify the effectiveness of perturbations generated the authors have formulated two objectives.

1. Fooling Objective : Ideally a perturbation should confuse the classifier so as to flip the benign prediction into a different adversarial prediction. In order to improve the efficacy of the perturbations, the author's use the confidence of the benign(Unperturbed/clean) prediction which should be reduced and that of another category should be made higher.

2. Diversity Objective: The idea is to encourage the generator to explore the space of perturbations and generate a diverse set of perturbations. This is done By increasing the distance between feature embeddings projected by the target classifier.

Setting up Data Manually

P.S: For the train split we randomly sampled 10 instances from each target class as described in the paper. Note: Dataset is Aproximately 7.5 GB in Size. Use Verify_Dataset.py to check for any errors.

For Unix Users:

• Run setup_dataset.bash to get the dataset. Automatically runs Verify_Dataset.py

For Windows Users:

• Option 1: Download from Archive.org

  • Archive Link

• Option 2 : Mega Download Link for Train abd Validation data of Imagenet 2012 (Obtained from Kaggle)

  • Validation Data: Mega Link
  • Training Data: Mega Link

• The Notebook Code.ipynb contains the code required for trainning the generator

• Notebooks in the directory contains utils and code to generate and interpolate perturbations from generator weights

• Note: Pretrained weights for Googlenet, Resnet50, VGG16 and VGG19 (Trained for 30 Epochs) can be found as a Kaggle Dataset Link : https://www.kaggle.com/gokkulnath/nag-pytorch-pretrained

Interpolating Latent Dimension for NAG :

Sample Generated Perturbations:

TODO :

• Add Additional Loss with Distance on Features ? Required?
• Fit, evaluate, predict ? , generate_perturbation
  • Function to Plot Fooling of a image across multiple classifiers
  • Read the paper and Documentation
  • Push to Cloud and save the model/host it
• Proof read the NAG Notebook and blog post
• Testing on clean images

Reference

@inproceedings{nag-cvpr-2018,
  title={NAG: Network for Adversary Generation},
  author={Mopuri, Konda Reddy and Ojha, Utkarsh and Garg, Utsav and Babu, R Venkatesh},
 booktitle = {Proceedings of the IEEE Computer Vision and Pattern Recognition ({CVPR})},
 year = {2018}
}