Research map / Adversarial robustness and security / Attacks on text and graphs
REAP: A Large-Scale Realistic Adversarial Patch Benchmark
Nabeel Hingun, Chawin Sitawarin, Jerry Li, David Wagner
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 |
|---|---|---|---|
| render patches on real signs with geometric and lighting transforms instead of on a synthetic sign | patch attack evaluation on Faster R-CNN (small patch) · REAP-S | ASR: 73.1 -> 39.2 | Hurt |
| render patches on real signs instead of synthetic sign | patch attack evaluation on YOLOF (small patch) · REAP-S | ASR: 88.6 -> 38.9 | Hurt |
| adversarial training of detectors | Faster R-CNN, large patch · REAP | FNR: 50.0 -> 13.9 | Improved |
| adversarial training of detectors | DINO, large patch · REAP | FNR: 62.8 -> 3.6 | Improved |
| adversarial training of detectors | clean performance · REAP | mAP (no patch): FRCNN 72.9 -> 73.3; YOLOF FNR 18.5 -> 21.0 | Mixed |
| remove the relighting transform when rendering patches (ablation) | YOLOF and Faster R-CNN · REAP | mAP: mAP decreases by 17 points for YOLOF and 15 for Faster R-CNN | Hurt |
| use full perspective transform instead of affine or translate-and-scale (ablation) | patch rendering · Realism test (44 samples) | RMSE: 1.72 -> 1.13 | Improved |
| use percentile relighting in RGB instead of HSV or LAB (ablation) | patch rendering · Realism test (44 samples) | RMSE: 0.227 -> 0.110 (HSV vs RGB) | Improved |
| use percentile relighting instead of polynomial fitting in RGB (ablation) | patch rendering · Realism test (44 samples) | RMSE: 0.113 -> 0.110 | No clear effect |
| per-instance attack instead of per-class patch against adversarially trained models (ablation) | Adv. Faster R-CNN · REAP-S small patch | ASR: 5.1 -> 16.0 | Improved |
| add color-jitter or uniform noise augmentation to attack EoT (ablation) | patch attack · REAP | FNR / mAP: FRCNN FNR 15.4 -> 16.1 (jitter 0.1) | No clear effect |
| transfer attack from another model instead of white-box per-class (ablation) | Adv. Faster R-CNN · REAP-S small patch | ASR: 5.1 -> 2.7 (from Synthetic) | Hurt |
Closest papers by meaning
- Revisiting Adversarial Patch Defenses on Object Detectors: Unified Evaluation, Large-Scale Dataset, and New Insights (ICCV 2025 · similarity 0.89)
- PAD: Patch-Agnostic Defense against Adversarial Patch Attacks (CVPR 2024 · similarity 0.87)
- Revisiting Adversarial Patches for Designing Camera-Agnostic Attacks against Person Detection (NeurIPS 2024 · similarity 0.87)
- SlowFormer: Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers (CVPR 2024 · similarity 0.86)
- PatchGuard: Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies (CVPR 2025 · similarity 0.86)
- Beware of Road Markings: A New Adversarial Patch Attack to Monocular Depth Estimation (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.
