Improved Diffusion-based Generative Model with Better Adversarial Robustness
Authors: Zekun Wang, Mingyang Yi, Shuchen Xue, Zhenguo Li, Ming Liu, Bing Qin, Zhi-Ming Ma
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, extensive empirical studies validate the effectiveness of AT in diffusion-based models. The code is available at https://github.com/kugwzk/AT_Diff. ... We conduct experiments with four diffusion samplers: IDDPM (Dhariwal & Nichol, 2021), DDIM (Song et al., 2022), DPM-Solver (Lu et al., 2022b), and ES (Ning et al., 2024) under various NFEs. The experimental results of CIFAR-10 and Image Net are shown in Table 1 and Table 2, respectively. |
| Researcher Affiliation | Collaboration | 1Harbin Institute of Technology 2Renmin University of China 3Academy of Mathematics and Systems Science, Chinese Academy of Sciences 4University of Chinese Academy of Sciences 5Huawei Noah s Ark Lab |
| Pseudocode | Yes | Algorithm 1 Adversarial Training for Diffusion Model ... Algorithm 2 Adversarial Training for Consistency Distillation |
| Open Source Code | Yes | Finally, extensive empirical studies validate the effectiveness of AT in diffusion-based models. The code is available at https://github.com/kugwzk/AT_Diff. |
| Open Datasets | Yes | We propose to conduct efficient AT on both DPM and CM in various tasks, including image generation on CIFAR10 32 32(Krizhevsky & Hinton, 2009) and Image Net 64 64 (Deng et al., 2009), and zero-shot Text-to-Image (T2I) generation on MS-COCO 512 512 (Lin et al., 2014b). |
| Dataset Splits | Yes | Following Luo et al. (2023) and Chen et al. (2024), we evaluate models on MS-COCO 2014 (Lin et al., 2014a) at a resolution of 512 512 by randomly drawing 30K prompts from its validation set. Then, we report the FID between the generated samples under these prompts and the reference samples from the full validation set following Saharia et al. (2022). |
| Hardware Specification | Yes | The models are trained in a cluster of NVIDIA Tesla V100s. |
| Software Dependencies | No | The paper mentions using the Adam W optimizer and mixed precision training but does not provide specific software library names with version numbers (e.g., PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | During training, we fine-tune the pretrained models (details are in Appendix E.1) with batch size 128 for 150K iterations under learning rate 1e-4 on CIFAR-10, and batch size 1024 for 50K iterations under learning rate of 3e-4 on Image Net. For the hyperparameters of AT, we select the adversarial learning rate α from {0.05, 0.1, 0.5} and the adversarial step K from {3, 5}. |