On Adversarial Training without Perturbing all Examples

Authors: Max Losch, Mohamed Omran, David Stutz, Mario Fritz, Bernt Schiele

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our subset analysis on a wide variety of image datasets like CIFAR-10, CIFAR-100, Image Net-200 and show transfer to SVHN, Oxford Flowers-102 and Caltech-256.
Researcher Affiliation Academia Max Losch1, Mohamed Omran1, David Stutz1, Mario Fritz2, Bernt Schiele1 1 Max Planck Institute for Informatics, Saarland Informatics Campus, 2 CISPA Helmholtz Center for Information Security, Saarbrücken {mlosch, mohomran, dstutz, schiele}@mpi-inf.mpg.de, fritz@cispa.de
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at http://github.com/mlosch/SAT.
Open Datasets Yes We evaluate our subset analysis on a wide variety of image datasets like CIFAR-10, CIFAR-100, Image Net-200 and show transfer to SVHN, Oxford Flowers-102 and Caltech-256. Image Net: a large-scale hierarchical image database. CVPR, 2009.
Dataset Splits Yes Throughout the course of the training, we evaluate AA after each learning rate decay on 10% of validation data Dval and perform a final evaluation with the model providing the highest robust accuracy. This evaluation is performed on the remaining 90% of validation data.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions software components like 'SGD' and 'PGD-7' and references a 'robustness (python library)' but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes For all training setups listed in table 2, we train our models from scratch using SGD with a momentum of 0.9. Adversarial training for the L2 norm is performed with 7 steps of projected gradient descent (PGD-7) within an ϵ2 = 0.5 for CIFAR and SVHN and ϵ = 3.0 for Image Net-200, Caltech-256 and Flowers-102. For each step, we use a step size of 0.1 and 0.5 respectively.