Skip to content

Tensorflow implementation for generating adversarial examples using convex programming

Notifications You must be signed in to change notification settings

ebalda/adversarialconvex

Repository files navigation

Generating Adversarial Examples using Convex Programming in tensorflow

This repository contains the tensorflow implementation of our paper: On Generation of Adversarial Examples using Convex Programming.

Citation:

 @INPROCEEDINGS{balda2018adversarialconvex,
   title={On Generation of Adversarial Examples using Convex Programming},
   author={Emilio Rafael Balda  and Arash Behboodi and Rudolf Mathar},
   booktitle={52th Asilomar Conference on Signals, Systems, and Computers},
   year={2018},
   month={Oct},
   organization={IEEE}
 } 

Dependencies:

  • Python 3
  • TensorFlow >= 1.4

Pre-trained models

For benchmarking we use the following pre-trained models and datasets:

  • MNIST dataset: A fully connected network with two hidden layers of size 150 and 100 respectively (test error = 1.7%), as well as the LeNet-5 architecture (test error = 0.9%).
  • CIFAR-10 dataset: The Network In Network (NIN) architecture (test error = 13.8%), and a 40 layer DenseNet (test error = 5.2%).

Usage

There are 3 main scripts: main.py, view_grahp.py, and get_example_images.py.

Visualize the graph of a pre-trained model

To create the visualization files (needed by tensorboard) for a pre-trained model, stored by default in ./pretrainedmodels/, use view_grahp.py. For example: python view_grahp.py --model2load=lenet

Compute the fooling ratio

To compute the fooling ratio using the pretrained models use main.py. For example:

python main.py --model2load=fcnn --n-images=1024 --max-epsilon=0.1
python main.py --model2load=lenet --n-images=1024 --max-epsilon=0.2
python main.py --model2load=nin --n-images=1024 --max-epsilon=0.03
python main.py --model2load=densenet --n-images=1024 --max-epsilon=0.01

result

These figures are stored in ./figures/ by default, the values that appear in these figrues are stored by default in ./results/ as csv files. To customize the loading/storing directories teake a look at

python main.py --help

Compute the robustness measures of a given classifier

For this use the --rho-mode option of main.py. For example:

python main.py --rho-mode --model2load=fcnn --n-images=2048
python main.py --rho-mode --model2load=lenet --n-images=2048
python main.py --rho-mode --model2load=nin --n-images=2048

or

python main.py --rho-mode --model2load=densenet --n-images=2048

The obtained values of the robustness metrics are stored in ./results/ by default.

Get example images

To add example images into ./examples/ you may use get_example_images.py. See python get_example_images.py --help for more information.

Questions?

Please drop me an e-mail if you have any questions!