They won't need to use counter techniques beyond fixing incorrect output from their models by making the general training methods more robust to features not seen by humans.