Research map / Adversarial robustness and security / Attacks on text and graphs
APRICOT: A Dataset of Physical Adversarial Attacks on Object Detection
A. Braunegg, Amartya Chakraborty, Michael Krumdick, Nicole Lape, Sara Leary, Keith Manville, Elizabeth Merkhofer, Laura Strickhart, Matthew Walmer
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 |
|---|---|---|---|
| move patches from digital insertion to physical photographs | FRCNN white-box patch · targeted fooling rate | fooling event rate: 0.50 -> 0.17 | Hurt |
| move from digital to physical | all patches · untargeted fooling rate on FRCNN | fooling event rate: 0.77 -> 0.38 | Hurt |
| increase patch size from small to large | FRCNN patch · FRCNN white-box digital targeted | fooling event rate: 0.07 -> 0.83 | Improved |
| use square instead of circle patches | FRCNN patch · FRCNN white-box physical targeted | fooling event rate: 0.23 -> 0.12 | Hurt |
| use square instead of circle patches | FRCNN patch · FRCNN white-box digital targeted | fooling event rate: 0.47 -> 0.53 | No clear effect |
| increase viewing angle from head-on to severe | FRCNN patch · FRCNN white-box physical targeted | fooling event rate: 0.21 -> 0.0 | Hurt |
| transfer patch to a different detector (black-box) | SSD patch on FRCNN vs SSD · digital targeted | fooling event rate: 0.23 -> 0.08 | Hurt |
| jointly train to detect patches and COCO objects (ablation) | patch detector trained on high confidence patches · APRICOT AP IoU 0.5:0.95 | AP: 0.540 -> 0.686 | Improved |
| jointly train to detect patches and COCO objects (ablation) | patch detector trained on Gaussian patches · APRICOT AP IoU 0.5 | AP: 0.478 -> 0.936 | Improved |
| train on high confidence adversarial patches instead of standard adversarial patches (ablation) | joint patch detector · APRICOT AP IoU 0.5:0.95 | AP: 0.670 -> 0.686 | Improved |
| train on adversarial patches instead of Gaussian noise patches (ablation) | joint patch detector · APRICOT AP IoU 0.5 | AP: 0.936 -> 0.937 | No clear effect |
| combine KDE and Bayesian uncertainty by logistic regression instead of uncertainty alone (ablation) | fooling-event flagging · APRICOT detections | ROC | No clear effect |
Closest papers by meaning
- Revisiting Adversarial Patches for Designing Camera-Agnostic Attacks against Person Detection (NeurIPS 2024 · similarity 0.89)
- Adversarial Attacks on Event-Based Pedestrian Detectors: A Physical Approach (AAAI 2025 · similarity 0.88)
- PAD: Patch-Agnostic Defense against Adversarial Patch Attacks (CVPR 2024 · similarity 0.87)
- DAP: A Dynamic Adversarial Patch for Evading Person Detectors (CVPR 2024 · similarity 0.87)
- GLOW: Global Layout Aware Attacks on Object Detection (CVPR 2024 · similarity 0.87)
- PatchGuard: Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies (CVPR 2025 · similarity 0.87)
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.
