Localized Randomized Smoothing for Collective Robustness Certification

Authors: Jan Schuchardt, Tom Wollschläger, Aleksandar Bojchevski, Stephan Günnemann

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we compare our method to all existing collective certificates for ℓp-norm perturbations: Center smoothing using isotropic Gaussian noise (Kumar & Goldstein, 2021), Seg Certify (Fischer et al., 2021) and the collective certificates of Schuchardt et al. (2021).
Researcher Affiliation Academia 1Technical University of Munich 2CISPA Helmholtz Center for Information Security
Pseudocode No The paper describes the proposed methods and derivations mathematically but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes An implementation will be made available at https://www.cs.cit.tum.de/daml/localized-smoothing.
Open Datasets Yes We evaluate our certificate for l2 perturbations on 100 images from the Pascal VOC (Everingham et al., 2010) 2012 segmentation validation set. Training is performed on 10582 samples extracted from SBD, also known as Pascal trainaug (Hariharan et al., 2011).
Dataset Splits Yes We evaluate our certificate for l2 perturbations on 100 images from the Pascal VOC (Everingham et al., 2010) 2012 segmentation validation set.
Hardware Specification Yes The experiments on Pascal-VOC with strictly local models (Fig. 2) were performed using a Xeon E5-2630 v4 CPU @ 2.20GHz, an NVIDA GTX 1080TI GPU and 128 GB of RAM. All other experiments were performed using an AMD EPYC 7543 CPU @ 2.80GHz, an NVIDA A100 GPU and 128 GB of RAM.
Software Dependencies Yes The collective linear program is solved using MOSEK (version 9.2.46) (MOSEK Ap S, 2019) through the CVXPY interface (version 1.1.13) (Diamond & Boyd, 2016).
Experiment Setup Yes We train our models for 512 epochs, using Dice loss and Adam(lr = 0.001, β1 = 0.9, β2 = 0.999, ϵ = 10-8, weight decay = 0). We use a batch size of 128 for Pascal-VOC and a batch size of 32 for Cityscapes.