Research map / Adversarial robustness and security / Attacks on text and graphs
All You Need Is RAW: Defending against Adversarial Attacks with Camera Image Pipelines
Yuxuan Zhang, Bo Dong, Felix Heide
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 |
|---|---|---|---|
| train F and G on real RAW (Zurich) instead of synthesized RAW (MIT) (ablation) | RAW-domain defense · ImageNet FGSM | Top-1 accuracy (%): 40.52 -> 58.85 | Improved |
| train on real RAW (Zurich) instead of synthesized RAW (MIT) (ablation) | RAW-domain defense · ImageNet C&W | Top-1 accuracy (%): 55.13 -> 71.40 | Improved |
| train on HDR-RAW-RGB instead of Zurich real RAW (ablation) | RAW-domain defense · ImageNet FGSM, C&W | Top-1 accuracy (%): 58.85 -> 55.57 (FGSM), 71.40 -> 71.65 (C&W) | Mixed |
| map through RAW instead of gamma-adjusted stage (ablation) | RAW-domain defense · ImageNet FGSM | Top-1 accuracy (%): 44.4 -> 57.33 | Improved |
| map through RAW instead of demosaic stage (ablation) | RAW-domain defense · ImageNet FGSM | Top-1 accuracy (%): 52.93 -> 57.33 | Improved |
| map through RAW instead of gamma stage (ablation) | RAW-domain defense · ImageNet PGD | Top-1 accuracy (%): 71% -> 53% | Improved |
| add G operator to the S operator (omega 0 to 0.7) (ablation) | RAW-domain defense · ImageNet FGSM | Top-1 accuracy (%): 64.25 -> 66.02 | Improved |
| add S operator to the G operator (omega 1 to 0.7) (ablation) | RAW-domain defense · ImageNet FGSM | Top-1 accuracy (%): 65.39 -> 66.02 | Improved |
| add Gaussian noise sigma 0.5 to F output (ablation) | RAW-domain defense · ImageNet FGSM | Top-1 accuracy (%): 66.02 -> 65.64 | No clear effect |
| mix F and G operators from different checkpoints (ablation) | RAW-domain defense · ImageNet | Top-1 accuracy (%): 66.02 -> 65.89 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ComDefend — ComDefend: An Efficient Image Compression Model to Defend Adversarial Examples (CVPR 2019) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Adversarial Imaging Pipelines (CVPR 2021 · similarity 0.89)
- Defense Against Adversarial Images Using Web-Scale Nearest-Neighbor Search (CVPR 2019 · similarity 0.88)
- SafetyNet: Detecting and Rejecting Adversarial Examples Robustly (ICCV 2017 · similarity 0.88)
- DISCO: Adversarial Defense with Local Implicit Functions (NeurIPS 2022 · similarity 0.88)
- One Man's Trash Is Another Man's Treasure: Resisting Adversarial Examples by Adversarial Examples (CVPR 2020 · similarity 0.88)
- Enhancing the Adversarial Robustness via Manifold Projection (AAAI 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.
