Globally-Robust Neural Networks

                    Klas Leino 1 Zifan Wang 1 Matt Fredrikson 1

             Abstract
   The threat of adversarial examples has motivated
   work on training certifiably robust neural networks
   to facilitate efficient verification of local robust-
   ness at inference time. We formalize a notion of
   global robustness, which captures the operational
   properties of on-line local robustness certification
   while yielding a natural learning objective for
   r ...