Closed paullintilhac closed 2 years ago
I still haven't done this yet, but it shouldn't be complicated. But after reading carlini's papers it seems hat the mean distortion is simply the metric you use for iterative attacks where the epsilon is not fixed, but is rather the result of some constrained minimization problem (i.e. minimizing the distortion such that it is misclassified).
There is code included at the bottom of the strategic_with_carlini
code that calculates the l2 distances for each adv example -- this can be used to easily calculate the mean distortion. So I will close this for now even though it isn't fully implemented yet.
figure out what carlini means by "mean distortion", calculate it for all attacks and defenses