QuadAttac$K$: A Quadratic Programming Approach to Learning Ordered Top-$K$ Adversarial Attacks

Authors: Thomas Paniagua, Ryan Grainger, Tianfu Wu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, the proposed Quad Attac K is tested in the Image Net-1k classification using Res Net-50, Dense Net-121, and Vision Transformers (Vi T-B and DEi T-S).4 Experiments
Researcher Affiliation Academia Thomas Paniagua, Ryan Grainger and Tianfu Wu Department of Electrical and Computer Engineering, NC State {tapaniag, rpgraing, twu19}@ncsu.edu
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Details are provided in the Appendix A and our released code repository.
Open Datasets Yes In this section, we evaluate our Quad Attac K with K = 1, 5, 10, 15, 20 in the Image Net-1k benchmark [Russakovsky et al., 2015] using two representative pretrained Conv Nets: the Res Net-50 [He et al., 2016] and the Dense Net-121 Huang et al. [2017], and two representative pretrained Transformers: the vanilla Vision Transformer (Base) [Dosovitskiy et al., 2020] and the data-efficient variant DEi T (small) [Touvron et al., 2021].
Dataset Splits Yes In Image Net-1k [Russakovsky et al., 2015], there are 50, 000 images for validation.
Hardware Specification No The paper does not specify the exact hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions "Py Torch" and "mmpretrain package" but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes The selection of learning rates γ and λ values (see Eqn. 10) in the context of learning attacks requires careful consideration to achieve the desired trade-offs (see Fig. 3) and optimize the attack performance. If K < 5, the learning rate is set to γ = 0.75e 3 (for all the four models). For our Quad Attac K... we set λ = 0.5 for Quad Attac K and λ = 5 for the logit/probability space losses for the K = 1 case. For all other values of K, we use λ = 10...