Research map / Detection and segmentation / CNN features
Detecting Backdoors in Object Detection via Pre-NMS Prediction Distribution Shift
Longtian Wang, Chenhao Lin, Zhengyu Zhao, Le Yang, Shiwei Wang, Yuhan Zhi, Xiaofei Xie, 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 |
|---|---|---|---|
| Use equal-size single-object crops as validation set (ablation) | DistScan (Faster R-CNN) · Val-set ablation, VOC, Faster R-CNN | avg detection accuracy (%): 67.59 -> 96.30 | Improved |
| Use equal-size crops instead of full images (ablation) | DistScan (YOLOv5) · Val-set ablation, VOC, YOLOv5 | avg detection accuracy (%): 100.00 -> 88.89 | Hurt |
| Enforce equal per-class counts vs random images (ablation) | DistScan (YOLOv5) · Val-set ablation, VOC, YOLOv5 | avg detection accuracy (%): 100.00 -> 100.00 | No clear effect |
| Apply confidence threshold 0.0005 vs none (ablation) | DistScan (Faster R-CNN) · Confidence threshold ablation, VOC, Faster R-CNN | detection accuracy (%): 50.00 -> 96.30 | Improved |
| Raise confidence threshold above 0.001 (ablation) | DistScan (Faster R-CNN) · Confidence threshold ablation, COCO, Faster R-CNN | detection accuracy (%): 95.37 -> 67.59 | Hurt |
| Raise confidence threshold to 0.01 for YOLOv5 on COCO (ablation) | DistScan (YOLOv5) · Confidence threshold ablation, COCO, YOLOv5 | detection accuracy (%): 96.30 -> 97.22 | No clear effect |
| DistScan vs ODSCAN on scene-level attacks | DistScan · Scene-level attacks (SPONGE, BLINDING, GMA) | detection accuracy (%): 50.00 -> 96.99 avg | Improved |
| DistScan vs MIA on GMA COCO Faster R-CNN | DistScan · GMA attack, COCO, Faster R-CNN | detection accuracy (%): 58.33 -> 100.00 | Improved |
| DistScan on additional attacks/architectures | DistScan · VOC extra attacks (YOLOv8, DETR, BadDet) | mean AUROC: 0.73 -> 0.87 | Improved |
| JS divergence vs KL, L2, cosine distance (ablation) | DistScan · Distance metric (appendix) | detection performance | Improved |
Closest papers by meaning
- Phantom: Physical Object Interactions as Dynamic Triggers for NMS-Exploited Backdoors (CVPR 2026 · similarity 0.90)
- Test-Time Backdoor Detection for Object Detection Models (CVPR 2025 · similarity 0.89)
- Backdoor Attacks on Open Vocabulary Object Detectors via Multi-Modal Prompt Tuning (AAAI 2026 · similarity 0.89)
- Lie Detector: Unified Backdoor Detection via Cross-Examination Framework (NeurIPS 2025 · similarity 0.86)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.86)
- A Closer Look at Backdoor Attacks on CLIP (ICML 2025 · 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.
