NeurIPS 2020

Certifying Confidence via Randomized Smoothing

Meta Review

Thank you for your submission to NeurIPS. While some of the reviewers felt that some aspects of this paper lacked much novelty (particularly, there was some debate about whether the technical results could be derived easily from existing results in the literature), there was also a consensus that the broad technique of using the full CDF of classifier scores in order to tighten randomized smoothing certification radii was an interesting one, and worthy of publication. I would recommend that in revising their paper, the authors strongly look at the connections to previous theory as mentioned by the reviewers, and determine whether the proofs they present can be more easily derived as consequences of past results (even as alternative derivations, if they feel the "self-contained" derivation is preferable to include). However, this doesn't diminish the overall value of the paper, which I still believe to be strong enough to warrant acceptance.