Diffusion-Based Adversarial Sample Generation for Improved Stealthiness and Controllability

Authors: Haotian Xue, Alexandre Araujo, Bin Hu, Yongxin Chen

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

Reproducibility Variable Result LLM Response
Research Type Experimental The experiment section aims to answer the following questions: (Q1) Is Diff-PGD/Diff-r PGD effective to generate adv-samples with higher realism? (Q2) Can Diff-PGD be easily applied to generate better style-customized adv-samples? (Q3) Can Diff-PGD be applied to physical world attacks? (Q4) Do adversarial samples generated by Diff-PGD show better properties like transferability and anti-purification ability? Datasets, Models, and Baselines. We use the validation dataset of Image Net [8] as our dataset to get some statistical results for global attacks and regional attacks.
Researcher Affiliation Academia Haotian Xue 1 Alexandre Araujo 2 Bin Hu 3 Yongxin Chen 1 1 Georgia Institute of Technology 2 New York University 3 University of Illinois Urbana-Champaign
Pseudocode Yes Algorithm 1 Diff-r PGD
Open Source Code Yes Code is available at https://github.com/xavihart/Diff-PGD
Open Datasets Yes We use the validation dataset of Image Net [8] as our dataset to get some statistical results for global attacks and regional attacks. [8] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248 255. Ieee, 2009.
Dataset Splits No The paper mentions using the "validation dataset of Image Net" and sampling 250 images from it, but it does not specify the train/validation/test split percentages or sample counts for the overall dataset used in experiments.
Hardware Specification Yes All the experiments conducted in this research were carried out on a single RTX-A6000 GPU, housed within a Ubuntu 20.04 server.
Software Dependencies No The paper states "implemented using the Py Torch framework" but does not specify any version numbers for PyTorch or other software dependencies.
Experiment Setup Yes We begin with the basic global ℓ digital attacks, where we set ℓ = 16/255 for PGD and Diff-PGD, and # of iterations n = 10 and step size η = 2/255. For Diff-PGD, we use DDIM with timestep Ts = 50 (noted as DDIM50 for simplicity), and K = 3 for the SDEdit module. Here we use ϵ = 16/255, η = 2/255, n = 10 as our major settings (except for the ablation study settings) for both PGD and Diff-PGD.