Research map / Adversarial robustness and security / Certified and empirical robustness
Part-Based Models Improve Adversarial Robustness
Chawin Sitawarin, Kornrapat Pongmala, Yizheng Chen, Nicholas Carlini, David Wagner
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 part-level segmentation supervision with the downsampled part model instead of omitting segmentation labels (ablation) | Downsampled Part Model · Part-ImageNet | clean accuracy: 76.9 -> 85.6 | Improved |
| Use part-level segmentation supervision with the downsampled part model instead of omitting segmentation labels (ablation) | Downsampled Part Model · Part-ImageNet | adversarial accuracy: 39.6 -> 39.4 | No clear effect |
| Use part-level segmentation supervision with the bounding-box part model instead of omitting segmentation labels (ablation) | Bounding-Box Part Model · Part-ImageNet | clean accuracy: 78.1 -> 86.5 | Improved |
| Use part-level segmentation supervision with the bounding-box part model instead of omitting segmentation labels (ablation) | Bounding-Box Part Model · Part-ImageNet | adversarial accuracy: 39.9 -> 39.2 | No clear effect |
| Replace segmentation masks with part bounding-box labels (ablation) | Part-based models · Part-ImageNet | clean accuracy: 85.6 -> 84.1 | Hurt |
| Replace segmentation masks with part centroid labels (ablation) | Part-based models · Part-ImageNet | clean accuracy: 85.6 -> 82.6 | Hurt |
| Replace segmentation masks with part bounding-box labels (ablation) | Part-based models · Cityscapes | clean accuracy: 94.8 -> 95.4 | Mixed |
| Replace segmentation masks with part bounding-box labels (ablation) | Part-based models · PASCAL-Part | clean accuracy: 77.3 -> 66.2 | Hurt |
| Use part segmentation supervision instead of object-level segmentation labels (ablation) | Part-based model · Part-ImageNet | accuracy | Improved |
| Reduce the fraction of samples with part segmentation annotations (ablation) | Part-based models · Part-ImageNet | adversarial accuracy | No clear effect |
| Use pseudo-labels from another segmentation model to reduce human segmentation labels (ablation) | Part-based models · Part-ImageNet | accuracy and robustness | No clear effect |
| Increase segmentation-loss weight (ablation) | Part-based models · Part-ImageNet | adversarial accuracy | Mixed |
Closest papers by meaning
- PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition (ECCV 2024 · similarity 0.93)
- Recognizing Object by Components With Human Prior Knowledge Enhances Adversarial Robustness of Deep Neural Networks (TPAMI 2023 · similarity 0.90)
- ImageNet-E: Benchmarking Neural Network Robustness via Attribute Editing (CVPR 2023 · similarity 0.87)
- RobustBench: a standardized adversarial robustness benchmark (NeurIPS 2021 · similarity 0.87)
- Evaluating the Adversarial Robustness of Semantic Segmentation: Trying Harder Pays Off (ECCV 2024 · similarity 0.87)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.86)
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.
