Comment by xvedejas

The fact that a human isn't fooled by the attack (we can still recognize the 32x32 images for what they are), points to an interesting gap in the abilities of conventional convolutional neural nets.