Research map / Adversarial robustness and security / Attacks on text and graphs
Revisiting Adversarial Patch Defenses on Object Detectors: Unified Evaluation, Large-Scale Dataset, and New Insights
Junhao Zheng, Jiahao Sun, Chenhao Lin, Zhengyu Zhao, Chen Ma, Chong Zhang, Cong Wang, Qian Wang, Chao Shen
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 |
|---|---|---|---|
| retrain defense on APDE dataset | SAC · APDE out-of-domain AdvCloak | person AP@0.5: 4.17 -> 71.29 | Improved |
| retrain defense on APDE dataset | Adyolo · APDE out-of-domain AdvCloak | person AP@0.5: 18.29 -> 22.36 | Improved |
| retrain defense on APDE dataset | NAPGuard · APDE out-of-domain AdvTshirt | person AP@0.5: 50.21 -> 70.89 | Improved |
| retrain defense on APDE dataset | SAC · T-SEA patches | person AP@0.5: 51.82 -> 71.61 | Improved |
| retrain defense on APDE dataset | existing defenses · APDE | AP@0.5: 15.09% AP@0.5 | Improved |
| use naturalistic patches (GNAP) instead of T-SEA as the attack | SAC · APDE failure cases | AP@0.5 after defense: 48.79 -> 73.47 | Hurt |
| use naturalistic patches (DM-NAP) instead of T-SEA as the attack | NutNet · APDE failure cases | AP@0.5 after defense: 68.03 -> 81.32 | Hurt |
| increase number of patches from 1 to 3 | SAC · T-SEA, YOLOv3 | person AP@0.5: 61.92 -> 57.33 | Hurt |
| increase number of patches from 1 to 3 | ObjectSeeker · T-SEA, YOLOv3 | person AP@0.5: 69.16 -> 68.91 (2 patches), / at 3 | Hurt |
| increase number of patches from 1 to 3 | NutNet · T-SEA, YOLOv3 | person AP@0.5: 87.42 -> 82.09 | Hurt |
| use empirical defense instead of certified (NutNet vs DetectorGuard) | NutNet · T-SEA, YOLOv3, 1 patch | person AP@0.5: 68.50 -> 87.42 | Improved |
| patch detection accuracy as selection criterion (NAPGuard best mIoU vs NutNet best AP) | defense ranking · APDE hiding attacks | mean person AP@0.5: 75.94 vs 76.53 | No clear effect |
Closest papers by meaning
- Segment and Complete: Defending Object Detectors Against Adversarial Patch Attacks With Robust Patch Detection (CVPR 2022 · similarity 0.95)
- Certified Defenses for Adversarial Patches (ICLR 2020 · similarity 0.93)
- PAD: Patch-Agnostic Defense against Adversarial Patch Attacks (CVPR 2024 · similarity 0.92)
- APRICOT: A Dataset of Physical Adversarial Attacks on Object Detection (ECCV 2020 · similarity 0.92)
- A Unified, Resilient, and Explainable Adversarial Patch Detector (CVPR 2025 · similarity 0.89)
- Revisiting Adversarial Patches for Designing Camera-Agnostic Attacks against Person Detection (NeurIPS 2024 · similarity 0.89)
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.
