Adversarial machine learning

This project explores adversarial machine learning. A convolutional neural network is used with a custom dataset of traffic signs to highlight the potential effects adversarial examples could have on self driving cars. The dataset consists of 2 classes "Stop" and "Go" ("Stop", "30mph"). Note that the README.md file in /dataset/traffic_sign/ has step by step instructions on how to create your own custom dataset for use with Tensorflow convolutional neural networks.

Recommendations

You'll need to have a firm understanding of the mechanics of convolutional neural networks to grasp the code. Check out the convolutional_neural_network wiki page next to the Settings tab. This will help explain the basics. At the bottom of the wiki page there are two very good YouTube videos, it's highly recommended you watch both.

Prerequisites

This project uses TensorFlow version 1.1.0. Follow the specific instructions on the link below for your machine (GPU support available for specific NVIDIA chipsets, if using a GPU you'll need NVIDIA CuDNN 5.1). Python 3.5 is also required along with various other modules including matplotlib for some of the tutorials. You should be guided through any missing modules when trying to run the code. It's important you use the correct version of TF and CuDNN !

Installation

Getting Started

git clone https://gitlab.ncl.ac.uk/securitylab/adversarial_ML.git

There are various example deep learning models in /tutorials/NN/. The various directories contain traditional .py and also .ipynb files for use with jupyter notebooks. These models use various datasets such as MNIST and cifar-10. They go from a very simple softmax classification with 92% accuracy on the MNIST data set, and later introduces ReLU activation, batch normalisation, dropout and convolutional neural networks. Eventually achieving an accuracy of 99.5% on MNIST with very good visualisations using the matplotlib module. The tutorials should work for the majority, however a couple may need tampering with to get them to work (due to different directory structures on your machine etc.)

Please note

• The various README.md files scattered around the repo may be slightly out of date as the project progresses. I will get around to fixing any issues however.
• Jupyter note books don't always play nicely with the Tensorboard suite of tools. So if you want to use .ipynb be aware of this.
• The /tutorials/GAN/ directory and the GAN wiki are there from research at the beginning of the project. I've left them in there but this project isn't using GANs.

Check the wiki for more documentation and background resources!