DiffAdvMAP: Flexible Diffusion-Based Framework for Generating Natural Unrestricted Adversarial Examples
Authors: Zhengzhao Pan, Hua Chen, Xiaogang Zhang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on Image Net show that we achieve a better trade-off between image quality, flexibility, and transferability than baseline unrestricted adversarial attack methods. |
| Researcher Affiliation | Academia | 1College of Electrical and Information Engineering, Hunan University, Changsha, China 2College of Computer Science and Electronic Engineering, Hunan University, Changsha, China. Correspondence to: Hua Chen <EMAIL>, Xiaogang Zhang <EMAIL>. |
| Pseudocode | Yes | C. Pseudo-code The pseudo-code of Diff Adv MAP and Diff Adv MAP-Region is shown in Alg.1 and Alg. 2 respectively. |
| Open Source Code | No | The text does not provide an explicit statement about the release of source code or a link to a code repository for the methodology described in this paper. |
| Open Datasets | Yes | We evaluate the performance of our framework on the Image Net-compatible dataset(Kurakin et al., 2018), consisting of 1,000 images from Image Net s validation set. |
| Dataset Splits | Yes | We evaluate the performance of our framework on the Image Net-compatible dataset(Kurakin et al., 2018), consisting of 1,000 images from Image Net s validation set. |
| Hardware Specification | Yes | All experiments are done with a single RTX3090 GPU. |
| Software Dependencies | No | The paper mentions various models and techniques used (e.g., latent diffusion model, DDPM, DDIM sampling, Inception V3, Resnet50) but does not provide specific version numbers for software dependencies such as programming languages, libraries, or frameworks. |
| Experiment Setup | Yes | Implementation Details. We leverage the DDIM sampling for the generation process. The number of diffusion steps T is respaced to 200, t = 40, c = 30 and the number of Diff Adv MAP iterations is set to I = 10 for generating UAEs from noise. For other attacking conditions, T = 100, t = 20, c = 40 and I = 2. We apply an adaptive learning rate with an initial value of lr = 0.01, ΞΎ i(i = 1, 2) in equation (10) is set to 0.1 for all settings. |