Implementing Adversarial Training in TensorFlow

Question

I would like to implement the following cost function for my neural network:

This makes use of adversarial inputs for the neural network to improve generalization [ref].

Specifically, I am having trouble with the part. In my TensorFlow graph, I have defined as an operation. How can I feed with an argument other than ?

The only way I have found to do this so far is to define a parallel network that shares the weights with my original network and passing it in its feed_dict argument. If possible, I would like to avoid having to redefine my entire network. How can I do this?

---

My TensorFlow model is written as:

x = tf.placeholder(tf.float32, [None, 32, 32]);
... # A simple neural network
y = tf.add(tf.matmul(h, W1), b1);
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y, t));

---

Possibly relevant:

tf.stop_gradient(input, name=None)

Stops gradient computation.

...lots more stuff...

• Adversarial training, where no backprop should happen through the adversarial example generation process.

https://www.tensorflow.org/versions/r0.7/api_docs/python/train.html#stop_gradient

Answer 1

You need to write your model in such a way that it supports calls like

output = model.fprop(input_tensor)

or

output = model.fprop(input_tensor, params)

The fprop method builds the same forward propagation expression twice, but with a different input tensor on each call:

raw_output = model.fprop(clean_examples)
adv_examples = ...
adv_output = model.fprop(adv_examples)

If you want to apply this to one of our open source models and it doesn't support the interface to do this, file an issue on github.