Research map / Image generation / Fast diffusion sampling
DiffAug: A Diffuse-and-Denoise Augmentation for Training Robust Classifiers
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 DiffAug, training classifiers on one-step diffuse-and-denoise augmentations from an unconditional ImageNet diffusion model, improving robustness (ImageNet-C, OOD, certified accuracy) and classifier-guided diffusion.
Problem
Improving classifier robustness to covariate shift, adversarial examples, and OOD inputs without extra data or expensive multi-step diffusion augmentation.
Approach
Augment training images by forward-diffusing to a random time and applying a single reverse-diffusion denoising step, treating the denoised outputs as label-preserving augmentations; also extends this to test-time ensembling (DE) and to denoising-augmented guidance classifiers.
Key results
- On ImageNet-C (severity 5) default mode, DAM+DiffAug improves average accuracy from 40.36 to 41.91 and RN50+DiffAug improves from 28.35 to 31.15; averaged over all evaluation modes, ViT-B/16 improves from 48.88 to 50.81 on ImageNet-C.
- AugMix+DiffAug with Scale reaches 84.81 AUROC on ImageNet near-OOD, reported as comparable to the top leaderboard entry (84.87).
- ViT+DiffAug certified accuracy at l2 radius 0.25 rises from 36.30% to 40.30% on 1k ImageNet samples; DA-classifier improves CIFAR10 guidance-classifier FID from 2.81 to 2.34.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DiffAug to classifier training | ResNet-50 and ViT-B/16 classifiers · ImageNet-C severity 5 | top-1 accuracy (%): RN50: 25.99 -> 29.44; ViT-B/16: 48.88 -> 50.81 | Helped |
| Add DiffAug to AugMix training | ResNet-50 with AugMix · ImageNet-C severity 5 | top-1 accuracy (%): 32.63 -> 35.33 | Helped |
| Add DiffAug to DeepAugment training | ResNet-50 with DeepAugment · ImageNet-C severity 5 | top-1 accuracy (%): 35.87 -> 38.28 | Helped |
| Add DiffAug to DeepAugment+AugMix training | ResNet-50 with DeepAugment+AugMix · ImageNet-C severity 5 | top-1 accuracy (%): 41.64 -> 43.57 | Helped |
| Add DiffAug to classifier training | ImageNet classifiers · ImageNet-Test | top-1 accuracy (%): RN50: 70.01 -> 71.73; ViT-B/16: 78.32 -> 79.52 | Helped |
| Add DiffAug to classifier training | ViT · ImageNet certified adversarial accuracy with DDS | certified accuracy (%): 36.30 -> 40.30; 25.50 -> 32.50; 16.72 -> 23.62; 14.10 -> 19.40; 10.70 -> 15.20; 8.10 -> 11.00 | Helped |
| Train with both noisy and denoised inputs | Guidance classifier · CIFAR10 diffused test examples | test accuracy: 54.79 -> 57.16 | Helped |
| Train with both noisy and denoised inputs | Guidance classifier · ImageNet diffused test examples | test accuracy: 33.78 -> 36.11 | Helped |
| Train with both noisy and denoised inputs | Guidance classifier · CIFAR10 image generation | FID: 2.81 -> 2.34 | Helped |
| Train with both noisy and denoised inputs | Guidance classifier · ImageNet image generation | FID: 5.44 -> 5.24 | Helped |
| Use DiffAug Ensemble test-time augmentation instead of DDA | ImageNet-C evaluation · ImageNet-C severity 5 | top-1 accuracy (%): 37.13 -> 39.63 | Helped |
| Use single-step DE rather than multi-step DDA | Test-time image adaptation · ImageNet-C inference | wallclock time per example (s): 4.75 -> 0.5 | Helped |
Surprising findings
- Training on partially synthesized (denoised) images does not degrade clean test accuracy, despite prior reports that synthetic data can hurt classification.
- A diffusion model trained only on ImageNet with no extra data improves robustness, contrary to prior findings that offline synthetic data from such models gives no benefit.
Limitations
- Single-step denoising is more expensive than handcrafted augmentations like AugMix, and DE does not match DDA on ImageNet-C for ViT.
- Evaluation pipelines for covariate shift, adversarial, and OOD robustness are separate, with no unified protocol.
- Classifier-guidance comparisons omit classifier-free guidance, and DA-classifier gradient computation is slower.
Experimental setup
Datasets: ImageNet-1k CIFAR10
Benchmarks: ImageNet-C ImageNet-R ImageNet-S ImageNet-Test OpenOOD near-OOD (SSB-Hard, NINCO) ImageNet certified accuracy (DDS) CIFAR10 and ImageNet generation (FID, IS, P/R/D/C)
Metrics: top-1 accuracy certified accuracy AUROC FPR95 FID IS precision recall density coverage
Baselines: AugMix DeepAugment DDA DDS OpenOOD ASH/ReAct/SCALE/MSP Noisy classifier
Builds on: Improved DDPM Diffusion Denoised Smoothing DDA AugMix DeepAugment DeiT-III
Closest papers by meaning
- Robust Classification via a Single Diffusion Model (ICML 2024 · similarity 0.87)
- DiffAug: Enhance Unsupervised Contrastive Learning with Domain-Knowledge-Free Diffusion-based Data Augmentation (ICML 2024 · similarity 0.87)
- Inversion Circle Interpolation: Diffusion-based Image Augmentation for Data-scarce Classification (CVPR 2025 · similarity 0.87)
- Efficient Image-to-Image Diffusion Classifier for Adversarial Robustness (AAAI 2025 · similarity 0.86)
- DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models (CVPR 2024 · similarity 0.86)
- Diffusion Models are Certifiably Robust Classifiers (NeurIPS 2024 · similarity 0.86)
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.