M4xim4l/InNOutRobustness

Hi~, Do you use the 500k-TI as In-distribution when using 500K as OOD?

Opened this issue · 1 comments

Hi~, Do you use the 500k-TI as In-distribution when using 500K as OOD?

You can use 500k-TI as in-distribution (in combination with the original CIFAR10 train set of course, the ratio is set to 0.7 as described in "Uncovering the Limits of Adversarial Training against Norm-Bounded Adversarial Examples" ) and then use all of 80M TI as out-distribution.