Research map / Adversarial robustness and security / Certified and empirical robustness
NAPPure: Adversarial Purification for Robust Image Classification under Non-Additive Perturbations
Junjie Nan, Jianing Li, Wei Chen, Mingkun Zhang, Xueqi Cheng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NAPPure instead of DiffPure | GTSRB-CNN · GTSRB average over attacks | robust accuracy (%): 47.61 -> 73.93 | Improved |
| use NAPPure instead of DiffPure | WideResNet · CIFAR-10 average over attacks | robust accuracy (%): 57.82 -> 68.55 | Improved |
| use NAPPure instead of DiffPure on large-scale data | ResNet-50 · ImageNet average over attacks (512 images) | robust accuracy (%): 29.62 -> 37.79 | Improved |
| use NAPPure instead of adversarial training under additive attack | GTSRB-CNN · GTSRB additive attack | robust accuracy (%): 47.85 -> 83.20 | Improved |
| use NAPPure instead of LM under additive attack | GTSRB-CNN · GTSRB additive attack | robust accuracy (%): 79.07 -> 83.20 | No clear effect |
| use type-agnostic NAPPure-joint instead of type-specific NAPPure | NAPPure · GTSRB average over attacks | robust accuracy (%): 73.93 -> 59.29 | Hurt |
| use NAPPure-joint interpolation instead of direct composition under composite attack (ablation) | NAPPure · GTSRB composite attack of all 4 perturbation types | robust accuracy (%): 37.10 -> 54.49 | Improved |
| increase purification iterations from 100 to 500 (ablation) | NAPPure · GTSRB patch attack | robust accuracy (%): 63.48% -> 74.22% | Improved |
| extend purification iterations from 500 to 1000 (ablation) | NAPPure · GTSRB patch attack | robust accuracy (%): 74.22% -> 60.74% | Hurt |
| replace 3-layer CNN auxiliary model with ResNet-18 (ablation) | NAPPure · GTSRB patch attack | robust accuracy (%): 74.22% -> 71.29% | No clear effect |
| enlarge attack patch size from 5x5 to 9x9 against fixed defense (ablation) | NAPPure · GTSRB patch attack | robust accuracy (%): 85.16% -> 67.97% | Hurt |
| enlarge blur attack kernel from 3x3 to 5x5 against 5x5 defense kernel (ablation) | NAPPure · GTSRB blur attack | robust accuracy (%): 91.80% -> 86.91% | Hurt |
Closest papers by meaning
- Instant Adversarial Purification with Adversarial Consistency Distillation (CVPR 2025 · similarity 0.88)
- Adversarial Training on Purification (AToP): Advancing Both Robustness and Generalization (ICLR 2024 · similarity 0.87)
- Diffusion-based Adversarial Purification from the Perspective of the Frequency Domain (ICML 2025 · similarity 0.87)
- CLIPure: Purification in Latent Space via CLIP for Adversarially Robust Zero-Shot Classification (ICLR 2025 · similarity 0.86)
- Towards Understanding the Robustness of Diffusion-Based Purification: A Stochastic Perspective (ICLR 2025 · similarity 0.86)
- DiffBreak: Is Diffusion-Based Purification Robust? (NeurIPS 2025 · similarity 0.85)
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.
