Scarhei / KerasAdversary

This project is aimed to provide a simple interface to attack the models built with Keras. It can be very simple to craft adversarial examples with different attacking methods using our API.

KerasAdversary

This library provides several simple but powerful interfaces on adversarial attacks.

Dependencies

This library is mainly designed to attack the target models built with Keras. However, RNN is not supported now.

To use our library in your project, several libraries are pre-requisites:

• Tensorflow (tested on 1.2.0)
• keras (tested on 2.0.8)
• scipy (to support visual functions in our library)
• matplotlib (to support visual functions in our library)
• numpy (to support matrix calculation)

Introductions

All the files can be found in the folder lib/.

KerasAdversary.py

This file mainly contains several classes for adversarial attacks.

1. WhiteBoxAttacks: L-BFGS-B, FGSM, BIM, C&W(TODO)
2. BlackBoxAttacks: TODO

Now it supports both target attacks and non-target attacks.

Adversarial attack can be quite simple with our library. For example, if model is the target model built with Keras, then it can be attacked by FGSM with the following commands:

from lib.AdversarialAttacks import WhiteBoxAttacks
import keras.backen as K	
AttackAgent = WhiteBoxAttacks(model, K.get_session())
adv_x = AttackAgent.fgsm(x, y, epsilon=0.1, clip_min=0., clip_max=1.)

where adv_x are adversarial examples of x.

NOTE: If you want to perform target attacks, then you should set the parameter target=True and y is your target label. For example,

from lib.AdversarialAttacks import WhiteBoxAttacks
import keras.backen as K	
AttackAgent = WhiteBoxAttacks(model, K.get_session())
adv_x = AttackAgent.fgsm(x, y, target=True, epsilon=0.1, clip_min=0., clip_max=1.)

visualization.py

This file contains several methods for visualizations.

Here are some examples:

1. mnist

   

2. cifar10

   

utils.py

This file contains several useful methods that you may use in your project.