Consistency Purification: Effective and Efficient Diffusion Purification towards Certified Robustness
Authors: Yiquan Li, Zhongzhu Chen, Kun Jin, Jiongxiao Wang, Jiachen Lei, Bo Li, Chaowei Xiao
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our comprehensive experiments demonstrate that our Consistency Purification framework achieves state-of-the-art certified robustness and efficiency compared to baseline methods. In this section, we begin by detailing the experimental settings, followed by our main results. Additionally, we conduct ablation studies to further demonstrate the effectiveness of our framework. All experiments are conducted with 1 NVIDIA RTX A5000 24GB GPU. |
| Researcher Affiliation | Academia | 1University of Wisconsin-Madison; 2 University of Michigan-Ann Arbor; 3California Institute of Technology ;4University of Illinois Urbana-Champaign |
| Pseudocode | Yes | We provide detailed descriptions of Consistency Purification in the following algorithms. Algorithm 1 presents the function of Consistency Fine-tuning and Consistency Purification respectively. Algorithm 2 shows the randomized smoothing algorithm from [19] with applying Consistency Purification to do prediction and compute the certified radius. |
| Open Source Code | Yes | We have submitted our code as the Supplementary Material. |
| Open Datasets | Yes | Dataset. We evaluate the Consistency Purification framework on both CIFAR-10 [35] and Image Net64 [36]. CIFAR-10 contains 32 32 pixel images across 10 different categories while Image Net-64 includes 64 64 pixel images across 1000 categories. |
| Dataset Splits | Yes | Due to limited computational resources, we select 500 test images for CIFAR-10 from the 10,000 CIFAR-10 test set, choosing every 20th example in sequence (e.g., the 1st, 21st, 41st, etc.). Similarly, for the Image Net-64 dataset, we sample 500 test examples from its 50,000 test examples using a fixed interval of 100. |
| Hardware Specification | Yes | All experiments are conducted with 1 NVIDIA RTX A5000 24GB GPU. |
| Software Dependencies | No | The paper does not specify version numbers for ancillary software dependencies. |
| Experiment Setup | Yes | Randomized Smoothing Settings. We set N = 10000 for both CIFAR-10 and Image Net as the number of sampling times used in randomized smoothing. We compute the certified radius for each test example at three different noise levels σ {0.25, 0.5, 1.0} for CIFAR-10 and σ {0.05, 0.15, 0.25} for Image Net-64. |