Overcoming the Convex Barrier for Simplex Inputs

Authors: Harkirat Singh Behl, M. Pawan Kumar, Philip Torr, Krishnamurthy Dvijotham

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We establish the scalability of our overall approach via the specification of ℓ1 robustness for CIFAR-10 and MNIST classification, where our approach improves the state of the art verified accuracy by up to 14.4%.
Researcher Affiliation Collaboration Harkirat Singh University of Oxford harkirat@robots.ox.ac.uk M. Pawan Kumar Deep Mind mpawan@deepmind.com Philip H.S. Torr University of Oxford phst@robots.ox.ac.uk Krishnamurthy (Dj) Dvijotham Deep Mind dvij@google.com
Pseudocode Yes Algorithm 1 Simplex Verify
Open Source Code No The paper states 'We used the publicly available training implementation of [Ding et al., 2019]...The code is made available under the LGPL License online 1. 1https://github.com/Borealis AI/advertorch.' This refers to a third-party library used for training, not the authors' own method.
Open Datasets Yes We evaluate the effectiveness of various methods for incomplete verification on the MNIST [Lecun] and CIFAR-10 [Krizhevsky and Hinton, 2009] datasets. ... The MNIST and CIFAR-10 datasets are widely used in the machine learning community, and are available under the creator s consent and MIT license respectively. ... For this experiment we verify the robustness of models on the UPMC FOOD-101 dataset [Wang et al., 2015] ... It is made available by the creators consent online 2. 2http://visiir.lip6.fr/
Dataset Splits No The paper mentions evaluating on test sets but does not specify training/validation/test dataset splits, exact percentages, or sample counts for reproduction.
Hardware Specification Yes Both the methods are run on 4 CPU threads on an Intel(R) Core(TM) i7-4960X CPU @ 3.60GHz processor. ... Both the Li RPA based solvers use Adam [Kingma and Ba, 2015] for updating the weighting vectors a, and are run on a single Nvidia Titan Xp GPU.
Software Dependencies No The paper mentions Gurobi and advertorch but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes The models are trained using the SLIDE attack (sparse ℓ1-descent attack) from Tramer and Boneh [2019] with ϵ = 0.3 for all networks except the VNN-comp big network, which is trained with ϵ = 0.05. ... We verify robustness against input perturbations lying in ℓ1 norm ball with ϵ = 0.35 for the MNIST network, ϵ = 0.2 for the VNN-comp big network and ϵ = 0.5 for all the other CIFAR-10 networks. ... All the methods use the same intermediate bounds, which are computed using Opt-Lirpa Planet run for 20 iterations.