First-Order Adversarial Vulnerability of Neural Networks and Input Dimension

Authors: Carl-Johann Simon-Gabriel, Yann Ollivier, Leon Bottou, Bernhard Schölkopf, David Lopez-Paz

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Section 4 empirically verifies the validity of the first-order Taylor approximation made in (2) and the correspondence between gradient regularization and adversarial augmentation (Fig.1). Section 4.2 analyzes the di-mension dependence of the average gradient-norms and adversarial vulnerability after usual and robust training.
Researcher Affiliation Collaboration 1Empirical Inference Department, Max Planck Institute for Intelligent Systems, Tübingen, Germany 2Facebook AI Research, Paris/New York.
Pseudocode No The paper describes methods and formulas but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Code available at https://github.com/facebookresearch/Adversarial And Dimensionality.
Open Datasets Yes We train several CNNs with same architecture to classify CIFAR-10 images (Krizhevsky, 2009).
Dataset Splits No The paper mentions training on CIFAR-10 images and evaluating on a 'held-out test-set', but does not explicitly describe a separate validation set split or its details.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for running experiments.
Software Dependencies No The paper mentions software like 'Foolbox-package' and 'scipy', but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes Section 4.1 uses an attack-threshold ϵ = 0.5% of the pixel-range (invisible to humans), with PGD-attacks from the Foolbox-package (Rauber et al., 2017). Section 4.2 uses self-coded PGD-attacks with random start with ϵ = 0.08%. As a safety-check, other attacks were tested as well (see App.4.1 & Fig.4), but results remained essentially unchanged. Note that the ϵ -thresholds should not be confused with the regularization-strengths ϵ appearing in (4) and (5), which will be varied. The datasets were normalized (σ .2). All regularization values ϵ are reported in these normalized units (i.e. multiply by .2 to compare with 0-1 pixel values). All nets had the same amount of parameters and very similar structure across input-resolutions (see Appendix G.1).