How to convince Inception you're a teapot.

Installation

You need Python (3.5 or above), Tensorflow, Keras and cleverhans to be installed.

Install cleverhans:

pip install -e git+https://github.com/tensorflow/cleverhans.git#egg=cleverhans

Install PIL:

pip install pillow

You may additionally need things like numpy - use:

pip install <thing>

for any scientific python libraries you are missing.

How to run

First, find an image you would like to use as an input and supply its path. In the example we attack an unsuspecting goldfish, but you can choose anything you like:

IMAGE_PATH=<path_to_your_image>

Now we're ready to go:

$ python teapot_attack.py

If the code crashes when it reaches the FastGradientSign method then you probably need to update your version of Tensorflow to be compatible with cleverhans.

Exercises

1. Find an image of a goldfish and run the iterative attack. What class is it classified after one iteration?

2. Try changing the parameter eps - what does it do? Does it change the answer to question 1?

3. Change the attack so that it is an untargeted attack. You'll need to change the arguements to the FastGradientSign method - check the documentation here. Just do this attack once don't repeat it - what might happen if you did repeat it? Tip:

fgsm_params = {'eps': eps, 'clip_min': -1.0, 'clip_max': 1.0}