Increasing Confidence in Adversarial Robustness Evaluations

Authors: Roland S. Zimmermann, Wieland Brendel, Florian Tramer, Nicholas Carlini

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For eleven out of thirteen previously-published defenses, the original evaluation of the defense fails our test, while stronger attacks that break these defenses pass it. We show that our test would have potentially identified eleven out of thirteen weak evaluations found in peer-reviewed papers.
Researcher Affiliation Collaboration Roland S. Zimmermann University of Tübingen Tübingen AI Center Wieland Brendel University of Tübingen Tübingen AI Center Florian Tramèr Google Nicholas Carlini Google
Pseudocode Yes Algorithm 1 Binarization Test for Classifiers with Linear Classification Readouts.
Open Source Code Yes Online version & code: zimmerrol.github.io/active-tests/ We included the implementation of our proposed test as well as the code to reproduce the results for the defenses investigated in this work.
Open Datasets Yes Since the dataset used in this work is a standard dataset (CIFAR-10) we do not discuss the aforementioned issues.
Dataset Splits Yes A detailed overview of the experimental details including the hyperparameters used is given in Section B
Hardware Specification Yes A description of the used hardware and total amount of compute is presented in section B.
Software Dependencies No The paper does not specify software dependencies with version numbers.
Experiment Setup Yes A detailed overview of the experimental details including the hyperparameters used is given in Section B