Adversarial Risk and the Dangers of Evaluating Against Weak Attacks

Authors: Jonathan Uesato, Brendan O’Donoghue, Pushmeet Kohli, Aaron Oord

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we empirically study several broad categories of proposed defense strategies on the CIFAR-10 (Krizhevsky and Hinton, 2009) and Image Net (Deng et al., 2009) datasets.
Researcher Affiliation Industry 1Deep Mind. Correspondence to: Jonathan Uesato <juesato@google.com>.
Pseudocode Yes Algorithm 1 SPSA adversarial attack
Open Source Code No The paper does not provide an unambiguous statement or a direct link to the source code for the methodology described in this paper. While 'cleverhans v2.1.0' is cited in the references and one of the authors is listed as an author of cleverhans, there is no explicit statement indicating that the code for *this paper's specific methods* is being released or is available.
Open Datasets Yes CIFAR-10 (Krizhevsky and Hinton, 2009) and Image Net (Deng et al., 2009) datasets.
Dataset Splits No The paper mentions evaluating on 'CIFAR-10 test set' and 'Image Net' and states that adversarial training achieves '47% accuracy on CIFAR-10 against a gradient-based adversary'. It refers to a 'held-out test set' but does not specify explicit training/validation splits (e.g., percentages or counts), nor does it cite specific predefined splits for training or validation portions beyond implying the use of standard test sets.
Hardware Specification No The paper mentions 'efficient GPU implementation' in the context of SPSA but does not provide any specific details about the hardware used for experiments, such as GPU models, CPU types, or memory specifications.
Software Dependencies Yes The paper lists 'cleverhans v2.1.0' in its bibliography, and one of the authors of this paper (Jonathan Uesato) is also an author of 'cleverhans v2.1.0: an adversarial machine learning library'. This indicates the use of this specific versioned library.
Experiment Setup Yes The paper provides specific values for experimental parameters, such as 'Perturbation sizes ϵ are relative to images with pixel intensities in [0, 255]', 'JPEG quality 75', and 'ϵdefend = 16'. It also mentions parameters for the SPSA algorithm: 'perturbation size δ, step size α > 0, batch size n'.