Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Examples for Semantic Segmentation and Object Detection
Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, Alan Yuille
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 |
|---|---|---|---|
| keep the spatial structure of the perturbation instead of randomly permuting it (ablation) | DAG on FR-ZF-07 · PascalVOC-2007 test | mAP: 58.33 -> 3.61 | Improved |
| keep the spatial structure of the perturbation instead of randomly permuting it (ablation) | DAG on FCN-VGG · VOC segmentation (736 val images) | mIoU: 65.47 -> 4.09 | Improved |
| keep the spatial structure of the perturbation instead of randomly permuting it (ablation) | DAG on FR-VGG-0712 · PascalVOC-2007 test | mAP: 71.97 -> 3.36 | Improved |
| use a denser proposal set (higher NMS IOU rate, chosen 0.90) (ablation) | DAG on FR-ZF-07 · PascalVOC-2007 test | mAP | Improved |
| train the victim detector on more data (VOC07+12 instead of VOC07) | DAG attack on FR-ZF · PascalVOC-2007 test | mAP: 58.70 -> 3.61 vs 61.07 -> 1.95 | Hurt |
| transfer perturbation across training sets (FR-ZF-07 to FR-ZF-0712) | DAG perturbation · PascalVOC-2007 test | mAP: 61.07 -> 22.15 | Improved |
| transfer perturbation across architectures (FR-ZF-07 to FR-VGG-07) | DAG perturbation · PascalVOC-2007 test | mAP: 69.14 -> 66.01 | No clear effect |
| transfer perturbation to black-box R-FCN-RN101 from FR-ZF-07 | DAG perturbation · PascalVOC-2007 test (black-box) | mAP: 78.06 -> 75.87 | No clear effect |
| attack DL-VGG with FCN-VGG perturbation (same VGG backbone) | DAG perturbation · VOC segmentation (736 val images, transfer) | mIoU: 70.72 -> 45.16 | Improved |
| cross-task attack with same architecture: FR-VGG-07 perturbation on FCN-VGG | DAG perturbation · VOC2012 seg val subset (687 images, cross-task) | mIoU: 54.87 -> 43.06 | Improved |
| add FR-VGG-0712 perturbation to FR-ZF-0712 perturbation (r2+r4) | DAG perturbation · PascalVOC-2007 test | mAP: 68.17 -> 4.28 | Improved |
| combine r1+r3+r5+r7 instead of r1 alone for black-box attack | DAG perturbation on R-FCN-RN101 · VOC2012 seg val subset (110 images, black-box) | mAP: 76.00 -> 64.52 | Improved |
Closest papers by meaning
- Characterizing Adversarial Examples Based on Spatial Consistency Information for Semantic Segmentation (ECCV 2018 · similarity 0.90)
- On the Robustness of Semantic Segmentation Models to Adversarial Attacks (CVPR 2018 · similarity 0.90)
- Universal Adversarial Perturbations Against Semantic Image Segmentation (ICCV 2017 · similarity 0.90)
- Generative Adversarial Perturbations (CVPR 2018 · similarity 0.89)
- Dynamic Divide-and-Conquer Adversarial Training for Robust Semantic Segmentation (ICCV 2021 · similarity 0.89)
- SemanticAdv: Generating Adversarial Examples via Attribute-conditioned Image Editing (ECCV 2020 · 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.
