English

Robust and Accurate -- Compositional Architectures for Randomized Smoothing

Machine Learning 2022-04-04 v1 Artificial Intelligence Cryptography and Security

Abstract

Randomized Smoothing (RS) is considered the state-of-the-art approach to obtain certifiably robust models for challenging tasks. However, current RS approaches drastically decrease standard accuracy on unperturbed data, severely limiting their real-world utility. To address this limitation, we propose a compositional architecture, ACES, which certifiably decides on a per-sample basis whether to use a smoothed model yielding predictions with guarantees or a more accurate standard model without guarantees. This, in contrast to prior approaches, enables both high standard accuracies and significant provable robustness. On challenging tasks such as ImageNet, we obtain, e.g., 80.0%80.0\% natural accuracy and 28.2%28.2\% certifiable accuracy against 2\ell_2 perturbations with r=1.0r=1.0. We release our code and models at https://github.com/eth-sri/aces.

Keywords

Cite

@article{arxiv.2204.00487,
  title  = {Robust and Accurate -- Compositional Architectures for Randomized Smoothing},
  author = {Miklós Z. Horváth and Mark Niklas Müller and Marc Fischer and Martin Vechev},
  journal= {arXiv preprint arXiv:2204.00487},
  year   = {2022}
}

Comments

Presented at the ICLR 2022 Workshop on Socially Responsible Machine Learning

R2 v1 2026-06-24T10:34:47.858Z