soarlab / aaqnn

Adversarial Attacks for Quantized Neural Networks

How to run

After you did both the setup of the QNN and for the attacker 'DeepGame' run the following:

python launcher.py 'seed_number' 'first_image' 'last_image' 'dataset'

Where:

seed_number: set the seed for the execution. Note: if the seed does not match any of the stored weights (previous training) the net is going to be trained from scratch.

First Image: index of the first image. On Cifar10, mnist max value is 9999.

Last Image: index of the last image. On Cifar10, mnist max value is 10000.

dataset: we support mnist, cifar10, fashion.

How to start the process on CloudLab

$ (screen -R)
$ git clone https://github.com/soarlab/AAQNN.git && cd AAQNN && ./repo-init.sh && source venv/bin/activate
$ python launcher.py 10 0 1 mnist 
$ (ctr+a d)
$ (exit)

How to run the analysis on super-computer

python launcher.py 10 0 10000 mnist

This is how we should run the analysis. What we need to tune is only the variable 'concurrentProcesses' described in the following.

'concurrentProcesses' describes how many pairs (ImageNumber, QuantizationLevel) are being attacked at the same time.

Ex. (img=10,Q=2) and (img=10,Q=4) are two different pairs!

If we figure out there is space for more parallelization (ex. CPU's usage is 60%), we can easily modify in 'executor.py' the variable 'concurrentProcesses'.

Prerequisite

Installed pylearn2

Install

python 3.6
pip3 install -r requirements.txt

Test installation

python test_installation.py

If nothing crashes for a minute, it's all right, the installation is successful.

Run

python3 launcher.py

aaqnn's Issues

Cleaning up the repo a bit

We should:

• Remove pycache folders from github. No need to have those version controlled
• Organize scripts so that they are not in the top level folder
• Write a short document describing what each folder contains. The structure is unclear at this point.
  Given that we are probably going to leverage this code for quite a bit, it would be good to clean it up somewhat.

Add FGSM attack

As DeepGame seems not to be reproducible/deterministic, a new approach should be added.

One suggestion is FGSM attack: https://arxiv.org/abs/1412.6572

Cleverhans library can be leveraged for this: https://github.com/tensorflow/cleverhans/blob/master/cleverhans/attacks/fast_gradient_method.py

Add CW attack

As DeepGame seems not to be reproducible/deterministic, a new approach should be added.

One suggestion is CW attack introduced in: https://arxiv.org/abs/1608.04644

Cleverhans library can be leveraged for this: https://github.com/tensorflow/cleverhans/blob/master/cleverhans/attacks/carlini_wagner_l2.py

Organize structure of the project

• Organize scripts so that they are not in the top level folder
• Write a short document describing what each folder contains. The structure is unclear at this point. Given that we are probably going to leverage this code for quite a bit, it would be good to clean it up somewhat.