Research map / Adversarial robustness and security / Attacks on text and graphs
Training on Foveated Images Improves Robustness to Adversarial Attacks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 with R-Blur instead of clean images | ResNet · Ecoset, Imagenet; APGD attacks | accuracy: 12-25 percentage points (pp) | Helped |
| Use R-Blur rather than R-Warp | biologically inspired defenses · Ecoset, Imagenet; adversarial attacks | accuracy: up to 33 pp | Helped |
| Use R-Blur rather than VOneBlock | biologically inspired defenses · Ecoset, Imagenet; adversarial attacks | accuracy: up to 15 pp | Helped |
| Use R-Blur rather than adversarial training | ResNet · Ecoset, Imagenet; common corruptions | accuracy: up to 7 pp higher accuracy on average | Helped |
| Use R-Blur rather than adversarial training | ResNet · Ecoset, Imagenet; white-box attacks | accuracy | Hurt |
| Add adaptive blur and desaturation, even without additive noise (ablation) | R-Blur · Ecoset-10; APGD, ∥δ∥∞=0.008 | accuracy: 11 point improvement | Helped |
| Add multiple fixation points instead of one center fixation (ablation) | R-Blur · Ecoset-10; APGD, ∥δ∥∞=0.008 | accuracy | Helped |
| Add adaptive color desaturation (ablation) | R-Blur · Ecoset-10; APGD, ∥δ∥∞=0.008 | accuracy | Helped |
| Add adaptive blur rather than non-adaptive blur (ablation) | R-Blur · Ecoset-10; APGD, ∥δ∥∞=0.008 | accuracy: 5 points | Helped |
| Use dynamic fixation rather than five predefined fixation points (ablation) | R-Blur · Ecoset-10; APGD | accuracy | No clear effect |
| Train with Gaussian noise rather than R-Blur (ablation) | ResNet · Imagenet; adversarial attacks | accuracy: 45-58% (8-10 points) lower accuracy | Hurt |
| Train R-Blur without noise rather than with noise (ablation) | R-Blur · Ecoset, Imagenet; certified accuracy | certified accuracy: more than 50% of the certified accuracy | Hurt |
Closest papers by meaning
- Biologically Inspired Mechanisms for Adversarial Robustness (NeurIPS 2020 · similarity 0.88)
- A Combination of Noise and Bilateral Filters Achieve Supralinear and Scalable Adversarial Robustness in CNNs (CVPR 2026 · similarity 0.88)
- Finding Biological Plausibility for Adversarially Robust Features via Metameric Tasks (ICLR 2022 · similarity 0.88)
- Towards Achieving Adversarial Robustness by Enforcing Feature Consistency Across Bit Planes (CVPR 2020 · similarity 0.87)
- Benign Overfitting in Adversarial Training for Vision Transformers (ICML 2026 · similarity 0.87)
- Learning Bregman Divergences with Application to Robustness (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.