KERAS-DCGAN

Example implementations of adversarial agents with keras.

Example 1 - generating artificial images

In this example we learn:

• how to generate believable looking MNIST digit images
• how to classify MNIST digits

The agents that learn this don't see the raw MNIST image data at all! Rather knowledge is transferred from an adversarial discriminator which is trying to spot the difference between fake generated images and the real thing.

We train three adversarial deep learning models - one for each agent.

Agents:

• A generator which takes a random signal + a digit class. The generator model tries to produce images that look real, get correctly classified by the classifier, and get a high score from the discriminator.
• A discriminator which is given images and has to decide if they are fake or not. The discriminator model is the only thing exposed to the real MNIST training images. Ideally after a long time the discriminator's loss rises.
• A classifier which is given generated images and learns a very standard MNIST multiclass classification model that outputs the digit class of the image (0 - 9).

Usage

Training:

python dcgan.py --mode train --batch-size <batch_size>

$ python dcgan.py --mode train --path ~/generated-images --batch_size 256

Image generation:

python dcgan.py --mode generate --batch-size <batch_size>

$ python dcgan.py --mode generate --batch-size 64 --nice` : top 5% images according to discriminator

$ python dcgan.py --mode generate --batch-size 128

Result

generated images :

train process :

Acknowledgments

Based on jacobgil's implementation of Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks.

High level changes from that repository:

• Adds a third agent which classifies generated digits
• Generator given an input class as well as random IV

Example 2 - Adversarial Cryptography

An implementation of Learning to Protect Communications with Adversarial Neural Cryptography using keras.

Essentially three parties are trying to carry out a very classic cryptography problem. Alice and Bob have a shared key and which to communicate a message P without Eve being able to reconstruct the message.

Diagram from: nlml.github.io

Model:

You can render a nice viz of the models' nodes using Tensorboard:

Training:

Notes:

I wasn't sure exactly how do do the custom loss function for alice using Keras.