tutorial: accuracy of naive model on adversarial examples seems high

[goodfeli]

I got something > .6. Not sure why this would be; I'd expect something like .99 for such a large eps.

[goodfeli]

oh, maybe because it's a convolutional model. all my MNIST experiments in the past have been with fully connected models.

[npapernot]

I tried removing the convolutions and it decreases adversarial accuracy

[goodfeli]

OK, this seems legit then. Interesting that convolution is so effective on MNIST.