Unable to generate l2 attack examples

[ericcheng09]

I have a pre-trained keras MNIST model and I want to use my model to generate adversarial examples.

Here is the code I change in class MNISTModel() in setup_mnist.py :
`

    self.num_channels = 1
    self.image_size = 28
    self.num_labels = 10
    K.set_session(session)
    model = load_model('MNIST_model1.h5')
    self.model = model

`

But when I ran the code, I got some examples that have all zeros.

Should I load the model like this?
Thank you

[msingh27]

Yes, I am also facing a similar situation. Did you find some workaround to this ericcheng09.
Loading the model trained using train_models.py also some outputs adversarial examples which are having all zeros.

Any suggestions?
Thanks

[Nicholas Carlini]

When it outputs all zeros, that's an indication it failed to construct an adversarial example. Most of the time, this means the keras model you are loading includes a softmax activation layer as the final layer, and you should remove this (or otherwise change the predict function to be tf.log(self.model(x)).

[msingh27]

Thanks for the feedback. It works for most of the examples now.

[kakacharles10]

Can I generate cifar-10 attack examples on Windows?

[Nicholas Carlini]

I've never tested that, but that seems like a different issue than this one.