FDA: Feature Disruptive Attack

Aditya Ganeshan, B.S. Vivek, and R. Venkatesh Babu

IEEE / CVF International Conference on Computer Vision 2019 (ICCV'19)

This repository contains code for our recently published adversarial attack FDA. Our work focuses on attacking and disrupting the features of adversarial samples, which make them useless for multiple tasks.

For more information read our paper here!

For a detailed Collab notebook Walkthrough checkout here!

The code is implemented in TensorFlow with the use of TF-slim and CleverHans.

Instruction

1. Fetch the required libraries:

git clone https://github.com/BardOfCodes/cleverhans.git

git clone https://github.com/BardOfCodes/models.git

2. Fetch the required dataset and other files. The test dataset is a subset of the ImageNet 2012 ILSVRC dataset established in NIPS 2017 (NeurIPS now) challenge dataset (based on cleverhans/examples/nips17_adversarial_competition/dataset/final_dataset.csv).

# Fetch the dataset from my dropbox!
cd assets
wget https://www.dropbox.com/s/mi1nnywrvagsalo/mini_imagenet.tar?dl=1
mv 'mini_imagenet.tar?dl=1' mini.tar
tar -xvf mini.tar

# fetch the models
cd models
./model_fetch.sh

3. Run the test configuration:

# if required, install the required python packages:
pip install -r requirements.txt

# add path to CleverHans and TF-Slim in pythonpath
export PYTHONPATH="$PWD/cleverhans:$PWD/models/research/:$PYTHONPATH"

# run the evaluation script.
python scripts/run_eval.py --config config/test_config.yaml

The model should show print the results and also log more information in the results folder.

Acknowledgement:

I would like to thank:

1. My many friends, mentors and colleagues from Video Analytics Labs for having multiple ad-hoc discussions on this work and giving valuable suggestions.

2. Preferred Networks Inc for providing the freedom to pursue personal research directions.

Note: The code for many other ablations in the paper can be found in this repository (warning: code is messy).