saving/loading CNN weights and losses to run different attack algorithms

[ephi...@yahoo.com]

Using mnist_tutorial_tf.py and/or cifar10_tutorial_tf.py, I would like to train the Cleverhans embedded CNN model on the legitimate dataset images and save and then load the model weights and losses. It appears to me and my colleagues that the training does not have to be executed each time we want to test a new adversarial attach algorithm. I tried setting save=True in untils_tf.py however this was ineffective. In the end, I would like to be able to run several different attack trials using the same trained CNN weights and losses. FYI, from the FGM equations it appears that the saved model parameters are needed to compute/generate the adversarial effect on the images. Please advise. Are there FLAGS to bypass training and only load previously computed parameters? Thank you. Arnold

[Nicolas Papernot]

Hi Arnold, 

Have you tried the new picklable model? https://github.com/tensorflow/cleverhans/blob/master/cleverhans_tutorials/mnist_tutorial_picklable.py

-N

On Wed, Nov 7, 2018 at 12:23 PM 'ephi...@yahoo.com' via cleverhans dev <cleverh...@googlegroups.com> wrote:

Using mnist_tutorial_tf.py and/or cifar10_tutorial_tf.py, I would like to train the Cleverhans embedded CNN model on the legitimate dataset images and save and then load the model weights and losses. It appears to me and my colleagues that the training does not have to be executed each time we want to test a new adversarial attach algorithm. I tried setting save=True in untils_tf.py however this was ineffective. In the end, I would like to be able to run several different attack trials using the same trained CNN weights and losses. FYI, from the FGM equations it appears that the saved model parameters are needed to compute/generate the adversarial effect on the images. Please advise. Are there FLAGS to bypass training and only load previously computed parameters? Thank you. Arnold

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-de...@googlegroups.com.
To post to this group, send email to cleverh...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cleverhans-dev/cfada36e-aa73-473e-afdc-cb034593d736%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[ephi...@yahoo.com]

Terrific. Thank you. I will look into it. It appears to be spot on. Am I correct in guessing that afterward I can use the code evaluate_pickled_model.py to load my saved/trained model and evaluate an attack algorithm of my choosing? That would be great.

Arnold

On Thursday, November 8, 2018 at 1:35:36 PM UTC-5, Nicolas Papernot wrote:

Hi Arnold, 

Have you tried the new picklable model? https://github.com/tensorflow/cleverhans/blob/master/cleverhans_tutorials/mnist_tutorial_picklable.py

-N

On Wed, Nov 7, 2018 at 12:23 PM 'ephi...@yahoo.com' via cleverhans dev <cleverhans-dev@googlegroups.com> wrote:

Using mnist_tutorial_tf.py and/or cifar10_tutorial_tf.py, I would like to train the Cleverhans embedded CNN model on the legitimate dataset images and save and then load the model weights and losses. It appears to me and my colleagues that the training does not have to be executed each time we want to test a new adversarial attach algorithm. I tried setting save=True in untils_tf.py however this was ineffective. In the end, I would like to be able to run several different attack trials using the same trained CNN weights and losses. FYI, from the FGM equations it appears that the saved model parameters are needed to compute/generate the adversarial effect on the images. Please advise. Are there FLAGS to bypass training and only load previously computed parameters? Thank you. Arnold

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.

To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-dev+unsubscribe@googlegroups.com.
To post to this group, send email to cleverhans-dev@googlegroups.com.

[Nicolas Papernot]

Yes, that should be possible.

Glad we could help.

-N

On Thu, Nov 8, 2018 at 11:32 AM 'ephi...@yahoo.com' via cleverhans dev <cleverh...@googlegroups.com> wrote:

Terrific. Thank you. I will look into it. It appears to be spot on. Am I correct in guessing that afterward I can use the code evaluate_pickled_model.py to load my saved/trained model and evaluate an attack algorithm of my choosing? That would be great.

Arnold

On Thursday, November 8, 2018 at 1:35:36 PM UTC-5, Nicolas Papernot wrote:

Hi Arnold, 

Have you tried the new picklable model? https://github.com/tensorflow/cleverhans/blob/master/cleverhans_tutorials/mnist_tutorial_picklable.py

-N

On Wed, Nov 7, 2018 at 12:23 PM 'ephi...@yahoo.com' via cleverhans dev <cleverh...@googlegroups.com> wrote:

Using mnist_tutorial_tf.py and/or cifar10_tutorial_tf.py, I would like to train the Cleverhans embedded CNN model on the legitimate dataset images and save and then load the model weights and losses. It appears to me and my colleagues that the training does not have to be executed each time we want to test a new adversarial attach algorithm. I tried setting save=True in untils_tf.py however this was ineffective. In the end, I would like to be able to run several different attack trials using the same trained CNN weights and losses. FYI, from the FGM equations it appears that the saved model parameters are needed to compute/generate the adversarial effect on the images. Please advise. Are there FLAGS to bypass training and only load previously computed parameters? Thank you. Arnold

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.

To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-de...@googlegroups.com.
To post to this group, send email to cleverh...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/cleverhans-dev/cfada36e-aa73-473e-afdc-cb034593d736%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

You received this message because you are subscribed to the Google Groups "cleverhans dev" group.

To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-de...@googlegroups.com.
To post to this group, send email to cleverh...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cleverhans-dev/518304e2-fa4a-41c7-ad48-1ec78bd332f1%40googlegroups.com.

[Ian Goodfellow]

Yeah, check out this tutorial: https://github.com/tensorflow/cleverhans/blob/master/cleverhans_tutorials/mnist_tutorial_picklable.py

It should point you to the evaluation script.

Let us know if you have any questions / feedback on the tutorial.

To view this discussion on the web visit https://groups.google.com/d/msgid/cleverhans-dev/CAEmUmGbRC1_Q06YG-VxeqApsa5xYSvQ6HR9OmTAe39QxO%3DYDWA%40mail.gmail.com.

[ephi...@yahoo.com]

Will do. Thank you for your expert help.

Arnold

On Thursday, November 8, 2018 at 4:03:25 PM UTC-5, Ian Goodfellow wrote:

Yeah, check out this tutorial: https://github.com/tensorflow/cleverhans/blob/master/cleverhans_tutorials/mnist_tutorial_picklable.py

It should point you to the evaluation script.

Let us know if you have any questions / feedback on the tutorial.

On Thu, Nov 8, 2018 at 11:41 AM Nicolas Papernot <nic...@papernot.fr> wrote:

Yes, that should be possible.

Glad we could help.

-N

On Thu, Nov 8, 2018 at 11:32 AM 'ephi...@yahoo.com' via cleverhans dev <cleverhans-dev@googlegroups.com> wrote:

Terrific. Thank you. I will look into it. It appears to be spot on. Am I correct in guessing that afterward I can use the code evaluate_pickled_model.py to load my saved/trained model and evaluate an attack algorithm of my choosing? That would be great.

Arnold

On Thursday, November 8, 2018 at 1:35:36 PM UTC-5, Nicolas Papernot wrote:

Hi Arnold, 

Have you tried the new picklable model? https://github.com/tensorflow/cleverhans/blob/master/cleverhans_tutorials/mnist_tutorial_picklable.py

-N

On Wed, Nov 7, 2018 at 12:23 PM 'ephi...@yahoo.com' via cleverhans dev <cleverhans-dev@googlegroups.com> wrote:

Using mnist_tutorial_tf.py and/or cifar10_tutorial_tf.py, I would like to train the Cleverhans embedded CNN model on the legitimate dataset images and save and then load the model weights and losses. It appears to me and my colleagues that the training does not have to be executed each time we want to test a new adversarial attach algorithm. I tried setting save=True in untils_tf.py however this was ineffective. In the end, I would like to be able to run several different attack trials using the same trained CNN weights and losses. FYI, from the FGM equations it appears that the saved model parameters are needed to compute/generate the adversarial effect on the images. Please advise. Are there FLAGS to bypass training and only load previously computed parameters? Thank you. Arnold

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.

To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-dev+unsubscribe@googlegroups.com.
To post to this group, send email to cleverhans-dev@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/cleverhans-dev/cfada36e-aa73-473e-afdc-cb034593d736%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.

To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-dev+unsubscribe@googlegroups.com.
To post to this group, send email to cleverhans-dev@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/cleverhans-dev/518304e2-fa4a-41c7-ad48-1ec78bd332f1%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to the Google Groups "cleverhans dev" group.

To unsubscribe from this group and stop receiving emails from it, send an email to cleverhans-dev+unsubscribe@googlegroups.com.
To post to this group, send email to cleverhans-dev@googlegroups.com.