Research map / Adversarial robustness and security / Attacks on text and graphs
Certified Defenses for Adversarial Patches
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 IBP certificates for patch attacks | Image classifiers · MNIST, 2×2 patch | certified accuracy: 91.6% | Helped |
| Account for preprocessing when attacking LGS | LGS defense · CIFAR, 5×5 patch | empirical adversarial accuracy: 22.11% -> 0.5% | Hurt |
| Account for preprocessing when attacking DW | DW defense · CIFAR, 5×5 patch | empirical adversarial accuracy: 59.3% -> 15.6% | Hurt |
| Use all patch locations during certificate training (ablation) | IBP patch defense · MNIST, 2×2 patch | certified accuracy: 87.7% -> 91.5% | Helped |
| Increase random training patches from 1 to 10 (ablation) | Random-patch certificate training · MNIST, 2×2 patch | certified accuracy: 82.9% -> 87.7% | Helped |
| Select patches with guided training instead of random training (ablation) | Certificate training with 10 patches · MNIST, 2×2 patch | certified accuracy: 87.7% -> 88.9% | Helped |
| Use a large model with guided training (ablation) | CIFAR patch defense · CIFAR, 5×5 patch | certified accuracy: 25.0% -> 29.6% | Helped |
| Use mixed natural and IBP loss (ablation) | IBP certificate training · Patch certification | certified accuracy | No clear effect |
| Merge the final linear weights with the label-margin operation (ablation) | IBP certificate training · Patch certification | certified accuracy | Helped |
| Train with an increasing epsilon schedule (ablation) | IBP certificate training · Patch certification | training stability | Helped |
| Replace all-patch training with bound pooling (ablation) | IBP patch defense · MNIST, 2×2 patch | certified accuracy: 91.6% -> 89.9% | Hurt |
| Train square patches and test on lines (ablation) | Largest square-patch-trained model · MNIST, 4 pixels | certified accuracy: 91.6% -> 92.5% | Helped |
Closest papers by meaning
- (De)Randomized Smoothing for Certifiable Defense against Patch Attacks (NeurIPS 2020 · similarity 0.93)
- PAD: Patch-Agnostic Defense against Adversarial Patch Attacks (CVPR 2024 · similarity 0.90)
- A Unified, Resilient, and Explainable Adversarial Patch Detector (CVPR 2025 · similarity 0.88)
- On the Scalability of Certified Adversarial Robustness with Generated Data (NeurIPS 2024 · similarity 0.87)
- Revisiting Adversarial Patches for Designing Camera-Agnostic Attacks against Person Detection (NeurIPS 2024 · similarity 0.87)
- PatchGuard: Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies (CVPR 2025 · 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.