Scalable Certified Segmentation via Randomized Smoothing

                  Marc Fischer 1 Maximilian Baader 1 Martin Vechev 1

             Abstract
    We present a new certification method for image
    and point cloud segmentation based on random-
    ized smoothing. The method leverages a novel
    scalable algorithm for prediction and certification        (a) Attacked image     (b) Ground truth segmentation
    that correctly accounts for multiple testing, nec-
    essary for ensuring stat ...