Research map / Adversarial robustness and security / Attacks on text and graphs
PBCAT: Patch-Based Composite Adversarial Training against Physically Realizable Attacks on Object Detection
Xiao Li, Yiming Zhu, Yifan Huang, Wei Zhang, Yingzhe He, Jie Shi, Xiaolin Hu
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 |
|---|---|---|---|
| add global imperceptible noise to patch adversarial training (ablation) | PBCAT · Inria AdvPatch | AP50: 35.4 -> 72.8 | Improved |
| add global imperceptible noise to patch adversarial training (ablation) | PBCAT · Synthetic AdvTexture | AP50: 1.6 -> 24.9 | Improved |
| add gradient-guided patch partition and selection (ablation) | PBCAT · Synthetic AdvTexture | AP50: 24.9 -> 63.3 | Improved |
| select top-50% sub-patches by gradient instead of random (ablation) | PBCAT · Synthetic AdvTexture | AP50: 14.2 -> 63.3 | Improved |
| use 64 sub-patches instead of 16 (ablation) | PBCAT · Synthetic AdvTexture | AP50: 50.8 -> 60.2 | Improved |
| use 64 sub-patches instead of pixel-level partition (ablation) | PBCAT · Synthetic AdvTexture | AP50: 20.4 -> 60.2 | Improved |
| enlarge sampled patch scale factor from 2sqrt2/10 to 4sqrt2/10 (ablation) | PBCAT · Inria AdvPatch (16 sub-patches) | AP50: 78.3 -> 63.4 | Hurt |
| raise top-k ratio from 50% to 70% (ablation) | PBCAT · Synthetic AdvTexture | AP50: 60.2 -> 19.0 | Hurt |
| lower top-k ratio from 50% to 30% (ablation) | PBCAT · Inria AdvPatch | AP50: 77.6 -> 51.3 | Hurt |
| use PBCAT instead of l_inf-AT (AdvOD) | PBCAT · Synthetic AdvTexture | AP50: 30.5 -> 60.2 | Improved |
| use PBCAT instead of l_inf-AT (AdvOD) on FCOS | PBCAT · Inria AdvPatch (FCOS) | AP50: 29.9 -> 58.0 | Improved |
| use PBCAT instead of non-AT preprocessing defenses | PBCAT · Inria AdvPatch | AP50: 57.1 -> 77.6 | Improved |
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.
- Jedi — Jedi: Entropy-Based Localization and Removal of Adversarial Patches (CVPR 2023) · 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
- Segment and Complete: Defending Object Detectors Against Adversarial Patch Attacks With Robust Patch Detection (CVPR 2022 · similarity 0.89)
- Revisiting Adversarial Patch Defenses on Object Detectors: Unified Evaluation, Large-Scale Dataset, and New Insights (ICCV 2025 · similarity 0.88)
- APRICOT: A Dataset of Physical Adversarial Attacks on Object Detection (ECCV 2020 · similarity 0.88)
- Revisiting Adversarial Patches for Designing Camera-Agnostic Attacks against Person Detection (NeurIPS 2024 · similarity 0.86)
- A Unified, Resilient, and Explainable Adversarial Patch Detector (CVPR 2025 · similarity 0.86)
- PAD: Patch-Agnostic Defense against Adversarial Patch Attacks (CVPR 2024 · similarity 0.85)
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.
