From what little I have read on the topic adversarial attacks are succesfull on neural networks because in order to train them succesfully we design them to operate on a linear regime as much as possible. This linearity compounded with the high dimensionality of the problem (images) is what makes easy to find small perturbations that make the output oscillate wildly. My point is that "attacking machine learning" maybe should be renamed "attacking neural networks"
Since this blog post was written, the same authors have published further research on transferability which is what really makes these hard to defend against. They even present a proof of concept task that is resistant to transferring (although it is not super practical): https://arxiv.org/abs/1704.03453
A note -- if you're linking to arXiv, it's better to link to the abstract (https://arxiv.org/abs/1704.03453) rather than directly to the PDF. From the abstract, one can easily click through to the PDF; not so the reverse. And the abstract allows one to do things like see different versions of the paper, search for other things by the same authors, etc. Thank you!
I wonder if one could introduce a secondary classifier which is immune (or more resistant) to this kind of attack as a fail safe. One idea that comes to mind is to back the neural net with a random forest, which I imagine would be very hard to trick with this kind of attack as a collection of independent (key) weak learners are trained on the data. To trick a random forest, you would have to trick the majority of the trees within it.
Here is a paper from Bosch in that direction, it uses a second network to classify examples as adversarials: https://arxiv.org/abs/1702.04267
Using a fail safe network is hard because adversarial examples usually have a high accuracy at a false class. So using an accuracy threshold in the main network wouldn't work. Using a network as described in the paper and then a different kind of classifier might be worth trying. But it has also been shown that adversarial examples can transfer to different kind of models (don't know if random forests have been tried as well).
8 comments
[ 3.2 ms ] story [ 18.0 ms ] threadIf it interests anyone, I wrote up a bit about this and other threats related to machine learning models, a few weeks ago: https://matt.life/papers/security_privacy_neural_networks.pd...
Using a fail safe network is hard because adversarial examples usually have a high accuracy at a false class. So using an accuracy threshold in the main network wouldn't work. Using a network as described in the paper and then a different kind of classifier might be worth trying. But it has also been shown that adversarial examples can transfer to different kind of models (don't know if random forests have been tried as well).