MNIST Adversarial Image Generator

Download Instructions

git clone https://github.com/arakhmat/mnist-adversarial-image-generator

Usage

If you running the script for the first time:

# Specify '--train' flag to train the model
python generator.py -t

For consequent runs, use:

python generator.py

You can specify the following arguments when running the script:

'-t', type=bool,  action='store_true', help='train the network'
'-e', type=bool,  action='store_true', help='run evaluation'
'-s', type=int,   default=20000,       help='number of training steps'
'-b', type=int,   default=64,          help='training batch size'
'-o', type=int,   default=2,           help='label to replace'
'-n', type=int,   default=6,           help='adversarial label'
'-m', type=int,   default=10,          help='number of images to modify'
'-l', type=float, default=0.001,       help='adversarial learning rate'
'-r', type=float, default=0.0001,      help='regularization lambda'
'-p', type=bool,  action='store_true', help='modify only a single pixel'

Example

In this example, 10 images of the digit 2 were modified to be classified as the digit 6.

Columns (left to right): Original Image, Delta, Adversarial Image

One Pixel Example

In this example, the images that initially predict 2 but are close to predicting 6 were modified to predict 6 by changing only a single pixel.

Columns (left to right): Original Image, Delta, Adversarial Image

Acknowledgments

• A Guide to TF Layers: Building a Convolutional Neural Network
• Breaking Linear Classifiers on ImageNet by Andrej Karpathy
• Explaining and Harnessing Adversarial Examples by Ian J. Goodfellow, Jonathon Shlens and Christian Szegedy
• One pixel attack for fooling deep neural networks by Jiawei Su, Danilo Vasconcellos Vargas and Sakurai Kouichi