Attack Deterministic Conditional Image Generative Models for Diverse and Controllable Generation
Authors: Tianyi Chu, Wei Xing, Jiafu Chen, Zhizhong Wang, Jiakai Sun, Lei Zhao, Haibo Chen, Huaizhong Lin
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted experiments on different tasks, such as image inpainting, style transfer, and super-resolution, using pre-trained models which have no diversity ability. [...] Extensive experiments demonstrate the effectiveness of our method. |
| Researcher Affiliation | Academia | Tianyi Chu1, Wei Xing1, Jiafu Chen1, Zhizhong Wang1, Jiakai Sun1, Lei Zhao1*, Haibo Chen,2 Huaizhong Lin1 1Zhejiang University, 2Nanjing University of Science and Technology {chutianyi, wxing, chenjiafu, endywon, csjk, cszhl, linhz}@zju.edu.cn, hbchen@njust.edu.cn |
| Pseudocode | Yes | Algorithm 1: Adversarial attack on deterministic generative model, given pre-trained generative model fθ and cross-model Vision-language representation model CLIP. |
| Open Source Code | No | The paper does not include an unambiguous statement that the authors are releasing their code or provide a link to a code repository. |
| Open Datasets | No | The paper mentions using SOTA pre-trained models like La Ma (Suvorov et al. 2022) and Sty Tr2 (Deng et al. 2021) for evaluation on tasks like image inpainting and style transfer, but does not explicitly name the specific datasets used for their experiments or provide access information for those datasets. |
| Dataset Splits | No | The paper does not provide specific details about train/validation/test dataset splits used for their experiments. Their method is a plug-in for pre-trained models, and while experiments are conducted, the data partitioning is not described. |
| Hardware Specification | Yes | all experiments were conducted on a single RTX 3090 GPU. |
| Software Dependencies | No | The paper mentions using models like CLIP (Radford et al. 2021), La Ma (Suvorov et al. 2022), and Sty Tr2 (Deng et al. 2021) but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | The attack step is set to 10 by default. For image inpainting (ill-posed task), we use ϵ [0.01, 0.05] with cmin = 0.09, cmax = 0.09. For style transfer (semi-ill-posed task), we use ϵ [0.1, 0.25] with cmin = 0.25, cmax = 0.25. As shown in Fig. 4, when apply detailed with ϵ = 0.005 and cmin = 0.01, cmax = 0.01 to the Swin IR (Liang et al. 2021) model... |