CSAW-HACK-ML

ECE GY 9163 Final Project

Anup Upasani (asu224)

Neeraja Narayanswamy (nn2108)

Priyanka Shishodia (ps4118)

Instructions

Step 1: Downloading Data

Go into the data folder herein and copy the link in the Google Drive Link text file or get the link from below.

Download all data as is and put it into the data folder

Link to data files : https://drive.google.com/drive/folders/13o2ybRJ1BkGUvfmQEeZqDo1kskyFywab?usp=sharing

Step 2: Checking Imports

The following imports are required.

keras, tensorflow, h5py, numpy, matplotlib, random, opencv-python, datetime, scipy, imageio

Note that sys, shutil, and math are required as well, but these should already be part of the python installation

Step 2.5 (Optional): Checking Code

To see how the repair.py code generates test images, run the following command using a command prompt inside the eval folder herein

cd eval/
python3 repair.py

Note that this will take only a few minutes to run at maximum

To see how the repair.py code generates GoodNet models, run the following command using a command prompt inside the eval folder herein

python3 repair.py init

Note that this will take 30+ minutes to run

To see how the repair.py code generates GoodNet models and sets up STRIP, run the following command using a command prompt inside the eval folder herein

python3 repair.py init complex

Note that this will take 4+ hours to run

Step 3: Evaluating Individual Images

Note: Please place any NEW images to test inside the eval folder herein (NOT inside eval/poisoned_images). Note that eval/poisoned_images contains pre-generated poisoned images to test the eval scripts on

To evaluate an image with any eval script, use the following syntax, where items in brackets are user inputs specified below. Run the command using a command prompt inside the eval folder herein

1. Pre-generated poisoned images:

    python3 [eval_script] poisoned_images/[image]
   
    Where,
    	[eval_script] can be the following options:
   
    		eval_sunglasses.py, eval_anonymous1.py, eval_anonymous2.py, eval_mtmt.py
   
    	(Which corresponds to the sunglasses, anonymous 1, anonymous 2, and multi-target multi-trigger networks, respectively)
   
    	[image] can be any of the filenames (with extension) inside the eval/poisoned images
    	
   
    For example,
   
    	python3 eval_sunglasses.py poisoned_images/poisonres_sunglasses.png
   

2. New images (clean or poisoned):

    python3 [eval_script] [image]
    	
    Where, 
    	[eval_script] can be the following options
   
    		eval_sunglasses.py, eval_anon1.py, eval_anon2.py, eval_mtmt.py
   
    	(Which correspond to the sunglasses, anonymous 1, anonymous 2, and multi-target multi-trigger networks, respectively)
    		
    	[image] can be any of the filenames (with extension) that were added to the eval folder