Research map / Adversarial robustness and security / Attacks on text and graphs
Towards Automated Testing and Robustification by Semantic Adversarial Data Generation
Rakshith Shetty, Mario Fritz, Bernt Schiele
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 convex hull constraint on appearance codes instead of free l-infinity manipulation (ablation) | SA-Rand-App · COCO | mAP@0.5: 45.8 -> 46.5 | Improved |
| sample guiding templates from hard instances instead of randomly | SA-Rand-App · UnRel | mAP@0.5: 39.2 -> 39.6 | Improved |
| jointly optimize position along with appearance | SA-App · COCO | mAP@0.5: 46.6 -> 46.7 | Improved |
| use optimized adversary instead of worst-case template replacement (ablation) | Base+WorstT · UnRel | mAP@0.5: 39.2 -> 39.6 | Improved |
| edit two objects instead of one | SA · COCO | mAP@0.5: +0.3 mAP | Improved |
| repeat adversarial generation four times against the updated model | SA#2 · UnRel | mAP@0.5: 39.0 -> 40.8 | Improved |
| combine semantic adversary with AutoAugment | SA#2 · COCO | mAP@0.5: 47.0 -> 47.8 | Improved |
| fine-tune on semantic adversarial data vs PSIS data | SA#2 x4 · COCO | mAP@0.5: 46.7 -> 47.1 | Improved |
| add semantic adversarial data (single object, 5 iterations) | Base+FT · VOC | mAP@0.5: 74.0 -> 75.0 | Improved |
| edit two objects instead of one | SA x5 · VOC | mAP@0.5: 75.0 -> 74.0 | Hurt |
| set objectness weight lambda to 1.0 instead of 0.5 (ablation) | SA-App · BDD100k | mAP@0.5: 50.8 -> 51.4 | Improved |
| optimize position in addition to appearance | SA-App · BDD100k | mAP@0.5: 51.4 -> 51.2 | Hurt |
Closest papers by meaning
- Robustness and Generalization via Generative Adversarial Training (ICCV 2021 · similarity 0.87)
- SemanticAdv: Generating Adversarial Examples via Attribute-conditioned Image Editing (ECCV 2020 · similarity 0.87)
- Unadversarial Examples: Designing Objects for Robust Vision (NeurIPS 2021 · similarity 0.86)
- Natural Adversarial Examples (CVPR 2021 · similarity 0.86)
- Learning to Generate Synthetic Data via Compositing (CVPR 2019 · similarity 0.86)
- Enhancing the Adversarial Robustness via Manifold Projection (AAAI 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.
