Research map / Image generation / Fast diffusion sampling
Consistency Purification: Effective and Efficient Diffusion Purification towards Certified Robustness
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Consistency Purification, using a distilled consistency model as a one-step diffusion purifier for randomized smoothing, further refined by LPIPS-based Consistency Fine-tuning, achieving higher certified accuracy than one-step DDPM purification on CIFAR-10 and ImageNet-64.
Problem
Diffusion purification for certified robustness via randomized smoothing faces a trade-off: one-step DDPM purification is efficient but yields off-manifold purified images, while multi-step methods (PF-ODE, DensePure, etc.) are costly.
Approach
Use a consistency model distilled from a pretrained EDM/PF-ODE as a one-step purifier that maps noisy inputs to on-manifold clean images, then fine-tune it with an LPIPS loss on noise levels used in randomized smoothing to improve semantic alignment with the original image.
Key results
- On CIFAR-10, Consistency Purification with Consistency Fine-tuning reaches 79.4% certified accuracy at epsilon=0.25 versus 73.6% for onestep-DDPM, with a single network evaluation.
- On ImageNet-64, Consistency Fine-tuning raises certified accuracy at epsilon=0.0 from 53.0% (onestep-DDPM) to 69.0%.
- Consistency Fine-tuning with LPIPS loss outperforms l1 and l2 losses in certified accuracy on CIFAR-10 (79.4 vs 76.4 and 77.0 at epsilon=0.25).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace onestep-DDPM purification with Consistency Purification | ViT-B/16 smoothed classifier · CIFAR-10 | certified accuracy at ε=0.25: 73.6 -> 77.2 | Helped |
| replace onestep-DDPM purification with Consistency Purification | ViT-B/16 smoothed classifier · ImageNet-64 | certified accuracy at ε=0.0: 53.0 -> 61.0 | Helped |
| add Consistency Fine-tuning with LPIPS loss | Consistency Purification with ViT-B/16 · CIFAR-10 | certified accuracy at ε=0.5: 59.8 -> 62.4 | Helped |
| add Consistency Fine-tuning | Consistency Purification with ViT-B/16 · ImageNet-64 | certified accuracy at ε=0.0: 61.0 -> 69.0 | Helped |
| use ℓ1 loss instead of LPIPS for fine-tuning (ablation) | Consistency Purification · CIFAR-10 | certified accuracy at ε=1.0: 35.4 -> 31.4 | Hurt |
| use ℓ2 loss instead of LPIPS for fine-tuning (ablation) | Consistency Purification · CIFAR-10 | certified accuracy at ε=0.25: 79.4 -> 77.0 | Hurt |
| sample fine-tuning noise levels continuously from [0,1] instead of {0.25, 0.5, 1.0} (ablation) | Consistency Purification · CIFAR-10 | certified accuracy at ε=0.25: 79.4 -> 76.2 | Hurt |
| add diffusion-model fine-tuning to classifier fine-tuning (ablation) | Consistency Purification · CIFAR-10 | certified accuracy at ε=0.5: 63.4 -> 64.8 | Helped |
| fine-tune the diffusion model instead of using no fine-tuning (ablation) | Consistency Purification · CIFAR-10 | certified accuracy at ε=0.5: 59.8 -> 62.4 | Helped |
Surprising findings
- Consistency Purification with one step is competitive with multi-step PF-ODE EDM on CIFAR-10 at small radii (90.4 vs 89.6 at epsilon=0.0).
Limitations
- Not evaluated on high-resolution images such as ImageNet 256x256 due to lack of public consistency checkpoints.
- Only 500 CIFAR-10 and 100 ImageNet-64 test images used for evaluation.
- Compute for ImageNet 256 consistency training is beyond available resources.
Experimental setup
Datasets: CIFAR-10 ImageNet-64
Benchmarks: CIFAR-10 ImageNet-64
Metrics: certified accuracy at l2 radius epsilon FID
Baselines: onestep-DDPM onestep-EDM PF-ODE EDM Diffusion Calibration
Builds on: randomized smoothing consistency models DDPM PF-ODE Diffusion Denoised Smoothing
Closest papers by meaning
- Truncated Consistency Models (ICLR 2025 · similarity 0.89)
- SCott: Accelerating Diffusion Models with Stochastic Consistency Distillation (AAAI 2025 · similarity 0.88)
- Consistency Diffusion Bridge Models (NeurIPS 2024 · similarity 0.88)
- High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.88)
- Robust Representation Consistency Model via Contrastive Denoising (ICLR 2025 · similarity 0.88)
- Towards Understanding the Robustness of Diffusion-Based Purification: A Stochastic Perspective (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.