Robust learning of halfspaces under log-concave marginals

Jane Lange, Arsen Vasilyan

We say that a classifier is \emph{adversarially robust} to perturbations of norm $r$ if, with high probability over a point $x$ drawn from the input distribution, there is no point…