fabiocarrara / features-adversarial-det

Detect adversarial images from intermediate features in distance space

Adversarial examples detection in features distance spaces

This repo contains code to reproduce the experiments presented in "Adversarial examples detection in features distance spaces". The code trains models for adversarial detection based on intermediate features of the attacked classifier embedded into dissimilarity spaces.

Requirements

The main requirements are:

• Python 3
• pytorch 0.4 + torchvision
• tensorflow 1.8 + cleverhans

and can be installed with:

pip3 install -r requirements.txt

You will also need the following datasets to replicate the experiments:

• ILSVRC'12 (ImageNet) TRAIN split
• NIPS 2017 Adversarial Competition DEV dataset

Steps to reproduce experiments

• Create the folder images/original in the project folder and put the NIPS DEV images in it
• Modify the IMAGENET variable in reproduce.sh to point to the folder containing the ILSVRC'12 dataset (the script will point to the $IMAGENET/train/ folder)
• Run reproduce.sh

./reproduce.sh

The reproduce.sh bash script runs all the steps needed to reproduce the experiments presented in the paper, that is:

1. Features extraction from ILSVRC'12 TRAIN dataset
2. Class centroid / medoid computation
3. Generation of adversarial examples
4. Training of multiple detectors
5. Reproducing ROC plots

features-adversarial-det's Issues

Hello, I have the following questions about this paper.

1.I wonder if only adversarial examples are needed when training the detector? Need an original image? Or do you need labels for adversarial examples?
2.Hello, I would like to know when ‘the Features extraction from ILSVRC'12 TRAIN dataset’, do you use all the training set images? How many images did you use, if not all the images in the training set?
3.I want to know what adversarial_stats.py is mainly used for？