ytsmiling / lmt Goto Github PK

Table 1 in your paper shows the performance of your work under different size of the perturbations. However, I'm confused how to restrict the size of perturbations especially under optimized based attack, i.e C&W attack. And it seems that there are no suggestions in your code. Could you please show how can I optimize adversarial examples under max_epsilon? Thanks a lot!

@ytsmiling
In the code pooling.py, you multiply 'factor' to l.

However, I think the Lipschitz constant is not supposed to be 1.
Could you explain it more explicitly or provide me a reference?

The example under description of evaluation under attacks calls train.py whereas it's method says to call evaluate.py.