TensorFlow/Adversarial - Revision history

Admin: /* Running */

2017-10-27T00:13:13Z

Running

← Older revision		Revision as of 00:13, 27 October 2017
Line 48:		Line 48:

	Running this model is slightly more complicated than running the adversarial crypto network.		Running this model is slightly more complicated than running the adversarial crypto network.

			The adversarial text network steps are as follows:
			* fetch data
			* generate vocab
			* generate training/validation/test data
			* pretrain language model
			* train classifier
			* evaluate classifier on test data

	===Get Vocabulary Data===		===Get Vocabulary Data===

Admin: /* Adversarial Text */

2017-10-27T00:06:37Z

Adversarial Text

← Older revision		Revision as of 00:06, 27 October 2017
Line 44:		Line 44:

	Link to code: https://github.com/tensorflow/models/tree/master/research/adversarial_text		Link to code: https://github.com/tensorflow/models/tree/master/research/adversarial_text

			==Running==

			Running this model is slightly more complicated than running the adversarial crypto network.

			===Get Vocabulary Data===

			Start by obtaining the data, which is an 80 MB tar file, and decompress it:

			<pre>
			$ wget http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz -O /tmp/imdb.tar.gz

			$ tar -xf /tmp/imdb.tar.gz -C /tmp

			$ du -hs /tmp/aclImdb
			487M /tmp/aclImdb
			</pre>

			===Build the Vocabulary===

			Use a Bazel job to build the vocabulary from the data:

			<pre>
			$ IMDB_DATA_DIR=/tmp/imdb

			$ bazel run data:gen_vocab -- \
			--output_dir=$IMDB_DATA_DIR \
			--dataset=imdb \
			--imdb_input_dir=/tmp/aclImdb \
			--lowercase=False
			</pre>

			This uses a build rule called <code>gen_vocab</code> located in <code>data/BUILD</code>:

			<pre>
			py_binary(
			name = "gen_vocab",
			srcs = ["gen_vocab.py"],
			deps = [
			":data_utils",
			":document_generators",
			# tensorflow dep,
			],
			)
			</pre>

			This build vocabulary step is, unfortunately, failing. See this Github issue (1917): https://github.com/tensorflow/models/issues/1917

	==Adversarial Image Network==		==Adversarial Image Network==

Admin: /* Adversarial Text */

2017-10-26T19:08:22Z

Adversarial Text

← Older revision		Revision as of 19:08, 26 October 2017
Line 40:		Line 40:

	==Adversarial Text==		==Adversarial Text==

			This trains a neural network model to detect the sentiment in IMDB text. This illustrates semi-supervised learning.

			Link to code: https://github.com/tensorflow/models/tree/master/research/adversarial_text

	==Adversarial Image Network==		==Adversarial Image Network==

Admin: /* The Model */

2017-10-26T17:39:41Z

The Model

← Older revision		Revision as of 17:39, 26 October 2017
Line 37:		Line 37:
	* Define a method that evaluates the networks as-is and prints the percent losses		* Define a method that evaluates the networks as-is and prints the percent losses
	* Define a method that trains the network for a specified number of iterations, stopping early if the network reaches its target losses		* Define a method that trains the network for a specified number of iterations, stopping early if the network reaches its target losses
	* Define a method that calls the training function (above), then re-trains Eve several more times		* Define a method that calls the training function (above), then re-trains Eve several more times from scratch

	==Adversarial Text==		==Adversarial Text==

Admin: /* Adversarial Crypto */

2017-10-26T17:39:21Z

Adversarial Crypto

← Older revision		Revision as of 17:39, 26 October 2017
Line 9:		Line 9:
	This adversarial crypto neural network attempts to learn how to protect communications using the adversarial architecture.		This adversarial crypto neural network attempts to learn how to protect communications using the adversarial architecture.

	~~Link to paper~~: "Learning to Protect Communications with Adversarial Neural Cryptography": https://arxiv.org/abs/1610.06918		Paper: "Learning to Protect Communications with Adversarial Neural Cryptography"

			Link to paper: https://arxiv.org/abs/1610.06918

	Link to code: https://github.com/tensorflow/models/tree/master/research/adversarial_crypto		Link to code: https://github.com/tensorflow/models/tree/master/research/adversarial_crypto

Admin: /* TensorFlow Examples of Adversarial Neural Networks */

2017-10-26T17:23:23Z

TensorFlow Examples of Adversarial Neural Networks

← Older revision		Revision as of 17:23, 26 October 2017
Line 3:		Line 3:
	Adversarial neural networks use an architecture consisting of two separate neural networks - one network attempts to learn how to accomplish a task, and another network attempts to differentiate between the output of the first network and the "real" output.		Adversarial neural networks use an architecture consisting of two separate neural networks - one network attempts to learn how to accomplish a task, and another network attempts to differentiate between the output of the first network and the "real" output.

	=TensorFlow Examples ~~of Adversarial Neural Networks~~=		=TensorFlow Adversarial Examples=

	==Adversarial Crypto==		==Adversarial Crypto==
Line 38:		Line 38:

	==Adversarial Text==		==Adversarial Text==

			==Adversarial Image Network==

	=Flags=		=Flags=

Admin: /* Adversarial Crypto */

2017-10-26T16:31:02Z

Adversarial Crypto

Show changes

Admin: /* Train and Evaluate Method */

2017-10-26T16:23:56Z

Train and Evaluate Method

@@ Line 414: / Line 414: @@
 </pre>
-====Train and Evaluate Method====
+====Training Methods====
 ==Adversarial Text==

Admin: /* Do Evaluation Method */

2017-10-26T16:16:13Z

Do Evaluation Method

← Older revision		Revision as of 16:16, 26 October 2017
Line 353:		Line 353:
	====Do Evaluation Method====		====Do Evaluation Method====

	We now come to the definition of the function ~~that actually does the evaluation,~~ <code>doeval()</code>.		We now come to the definition of the function <code>doeval()</code> . This function evaluates how good the cryptosystem the Alice/Bob/Eve networks have created are at minimizing Eve's ability to decrypt the message with the ciphertext only and maximizing Bob's ability to decrypt the message with a ciphertext and key.

			This method evaluates only, and does not train. It runs the networks on a batch of n messages, computes the percentage of bits lost by Bob and Eve, prints the percentages, and returns them. These values are used to determine when to stop training the networks.

	The method header takes a few arguments:		The method header takes a few arguments:
Line 377:		Line 379:
	The main role of the doeval function is to run the neural network, and compute the losses that result. The TensorFlow session variable <code>s</code> will contain all three neural networks on its graph, so we can just call <code>s.run()</code> without needing to specify all three graphs.		The main role of the doeval function is to run the neural network, and compute the losses that result. The TensorFlow session variable <code>s</code> will contain all three neural networks on its graph, so we can just call <code>s.run()</code> without needing to specify all three graphs.

	~~Note that this~~ is ~~called in batches~~, ~~meaning~~ the ~~loss_accum variables~~ are ~~only computed ''within'' a batch and are reset each batch~~.		Link to Session class documentation: https://www.tensorflow.org/api_docs/python/tf/Session

			Link to Session.run documentation: https://www.tensorflow.org/api_docs/python/tf/Session#run

			The run method is passed a list of loss functions owned by the AdversarialCrypto class - these were the sums of incorrect bits defined above (<code>self.bob_reconstruction_loss = tf.reduce_sum(self.bob_bits_wrong)</code>).

			The reduce_sum function simply sums tensor components along a particular axis, thus reducing the dimensionality of the tensor. In this case we are summing the incorrect bits along the "axis" of the message.

	For each iteration <code>for _ in xrange(n)</code> we run the model, and get the loss functions from Bob and Eve. Recall from above that bob_reconstruction_loss and eve_loss are both the results of calls to ~~<code>tf.~~reduce_sum~~()</code>, which are computed on variables inside the model. The [~~https://www.tensorflow.org/api_docs/python/tf/reduce_sum reduce_sum] function simply sums tensor components along a particular axis, thus reducing the dimensionality of the tensor. In this case we are summing the incorrect bits along the "axis" of the message.		Link to reduce_sum documentation: https://www.tensorflow.org/api_docs/python/tf/reduce_sum

	~~At each iteration~~, we ~~are accumulating to total losses~~, ~~that~~ is, the ~~sum~~ of the bits ~~that were incorrect~~. ~~We accumulate these~~ for ~~each iteration~~.		When we call <code>s.run()</code>, we can pass a single graph element, or an arbitrarily nested collection containing graph elements at the leaves. The graph element can be an operation, a tensor, a tensor handle, or a string naming a tensor or operation in the graph; in this case, we pass a list of two tf.reduce_sum operations.

			Because the <code>s.run()</code> method is passed two values, it will return two values - the corresponding values returned from the two operations. Thus, the variable bl contains the result of the tf.reduce_sum call for Bob's network (the number of incorrect bits from Bob's network), while el contains the result of the tf.reduce_sum call for Eve's network.

	<pre>		<pre>
Line 392:		Line 402:
	</pre>		</pre>

	In the end, ~~we compute~~ the loss ~~percentage by multiplying~~		These error values are set to zero for each batch, and are accumulated across the batch.

			At the end of the batch of runs, the loss percentages are computed and returned.

	<pre>		<pre>

Admin: /* Do Evaluation Method */

2017-10-26T15:55:41Z

Do Evaluation Method

← Older revision		Revision as of 15:55, 26 October 2017
Line 375:		Line 375:
	</pre>		</pre>

	The main role of the doeval function is to run the neural network, and compute the losses that result. The TensorFlow session variable s will contain all three neural networks on its graph, so we can just call run without needing to specify all three graphs.		The main role of the doeval function is to run the neural network, and compute the losses that result. The TensorFlow session variable <code>s</code> will contain all three neural networks on its graph, so we can just call <code>s.run()</code> without needing to specify all three graphs.

	Note that this is called in batches, meaning the loss_accum variables are only computed ''within'' a batch and are reset each batch.		Note that this is called in batches, meaning the loss_accum variables are only computed ''within'' a batch and are reset each batch.