Learning Universal Adversarial Perturbation by Adversarial Example

Authors: Maosen Li, Yanhua Yang, Kun Wei, Xu Yang, Heng Huang1350-1358

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that the proposed method improves the performance by a significant margin over the existing methods in both data-dependent and data-independent settings.
Researcher Affiliation Academia Maosen Li1, Yanhua Yang1 , Kun Wei1, Xu Yang1, Heng Huang2 1Xidian University, Xi an 710071, China 2Department of Electrical and Computer Engineering, University of Pittsburgh, PA 15260, USA
Pseudocode Yes Algorithm 1: Our UAP algorithm
Open Source Code Yes Code is available at https://github.com/lisenxd/AT-UAP.
Open Datasets Yes We use the Image Net validation set (Russakovsky et al. 2015) containing 50,000 samples to evaluate the performance. We also explore generating data-dependent UAP with the Image Net training data.
Dataset Splits Yes We use the Image Net validation set (Russakovsky et al. 2015) containing 50,000 samples to evaluate the performance.
Hardware Specification Yes All of our experiments are conducted on Pytorch and run with single NVIDIA TITAN Xp GPU.
Software Dependencies No The paper mentions 'Pytorch' but does not provide a specific version number for this or any other software dependency.
Experiment Setup Yes The number of iterations T, batch-size m, learning rate γ and trade-off factor λ are set to 1000, 32, 0.5 and 0.05, respectively. ϵ2, constant a0 and zoom factor α are set to 20, 14, 20 respectively for data-independent setting and in data-dependent setting, ϵ2 is set to 4.