DISCO: Adversarial Defense with Local Implicit Functions

Neural Information Processing Systems 

More discussion can be found in Sec. In this section, we discuss the qualitative results of DISCO transferability across attacks. More discussion can be found in Sec. By default, we use s = 3 in all our experiments. For a single image Cifar10 of size 32x32, STL requires an Cifar10 5.9 For a single ImageNet image of size 224, STL requires 23.71 seconds while DISCO (K=1) only This shows that DISCO is a better defense in the sense that it can handle widely varying input image sizes with minor variations of computing cost.