TASK 4, COMP 551:: Adversarial Machine Learning

All source code are put under src subdirectory.

Project report file

adversarial-attacks-image.pdf

Link to the video

Link

Tools used, and other dependencies

1. Cleverhans, OpenAI
2. Tensorflow, version 3.4 and above
3. Keras

Generated adversarial examples

https://drive.google.com/drive/folders/0B2pDcGINKNIOVmdBc1Z4VlBGVGs?usp=sharing

Attack Files

attack_fgsm.py
attack_jsma.py
attack_blackbox.py

Defense Files

def_adv.py

Defensive Distillation

simple_NN_laddened_with_defensive_distillation.py
Contains the implementation of a protection of a network with defensive distillation.
Toggle comment between lines 74 and 75 to see performance on legitimate and adversarial examples. simple_NN_with_no_defensive_distillation.py
Contains a model with no protection from any threat.
Toggle comment between lines 50 and 51 to see performance on legitimate and adversarial examples.

Data Visualization and Results Generation Files

visualize.py
images.pysimple_NN_with_no_defensive_distillation.py

Comparison Between Different Methods

results.py

Other directions

1. All the files can be run using "python "
2. Depending on the location of datafile, the path has to be edited inside the file.
   Due to size limitations, adversarial example files are uploaded on google drive. Link has been shared at the top of this file.