Extensions and Limitations of Randomized Smoothing for Robustness Guarantees

Jamie Hayes; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 786-787

Abstract


Randomized smoothing, a method to certify a classifier's decision on an input is invariant under adversarial noise, offers attractive advantages over other certification methods. It operates in a black-box and so certification is not constrained by the size of the classifier's architecture. Here, we extend the work of Li et al. (2019), studying how the choice of divergence between smoothing measures affects the final robustness guarantee, and how the choice of smoothing measure itself can lead to guarantees in differing threat models. To this end, we develop a method to certify robustness against any Lp norm minimized adversarial perturbation. We then demonstrate a negative result, that randomized smoothing suffers from the curse of dimensionality; as p increases, the effective radius around an input one can certify vanishes.

Related Material


[pdf] [supp] [video]
[bibtex]
@InProceedings{Hayes_2020_CVPR_Workshops,
author = {Hayes, Jamie},
title = {Extensions and Limitations of Randomized Smoothing for Robustness Guarantees},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2020}
}