Research map / Adversarial robustness and security / Attacks on text and graphs
Recognizing Object by Components With Human Prior Knowledge Enhances Adversarial Robustness of Deep Neural Networks
Xiao Li, Ziqi Wang, Bo Zhang, Fuchun Sun, Xiaolin Hu
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 part segmentation supervision (w/ P, w/o K vs w/o P) (ablation) | ROCK · Pascal-Part-C PGD-40 eps=1 | accuracy: 14.4 -> 19.4 | Improved |
| add human prior knowledge linkage rules on top of part segmentation (ablation) | ROCK · Pascal-Part-C PGD-40 eps=1 | accuracy: 19.4 -> 27.5 | Improved |
| add importance weights of linkage rules (ablation) | ROCK · Pascal-Part-C PGD-40 eps=1 | accuracy: 27.5 -> 28.2 | Improved |
| add human prior knowledge (benign accuracy) (ablation) | ROCK · Pascal-Part-C benign | accuracy: 84.8 -> 80.7 | Hurt |
| replace ROW-50 with ROCK under white-box PGD-40 | ROW-50 · Pascal-Part-C eps=1 | accuracy: 3.5 -> 28.2 | Improved |
| replace ROW_X-101 with ROCK under SPSA black-box | ROW_X-101 · Pascal-Part-C eps=8 | accuracy: 9.8 -> 52.3 | Improved |
| combine with TRADES-AWP-EMA adversarial training | ROCK vs ROW-50 · PartImageNet-C AutoAttack eps=8 | accuracy: 11.5 -> 17.3 | Improved |
| use ResNeXt-101 backbone for ROCK with TRADES-AWP-EMA | ROCK · Pascal-Part-C PGD-40 eps=8 | accuracy: 32.6 -> 34.5 | Improved |
| use DeiT-Ti transformer backbone in ROCK | DeiT-Ti · Pascal-Part-C PGD-40 eps=1 | accuracy: 15.8 -> 27.0 | Improved |
| use only 10 part-labeled images per category instead of all part labels (ablation) | ROCK · PartImageNet-C 50 categories PGD-40 eps=1 | accuracy: 19.4 -> 14.5 | Hurt |
| replace real part labels with unsupervised pseudo part labels (ablation) | ROCK · PartImageNet-C 45 categories PGD-40 eps=1 | accuracy: 22.6 -> 19.6 | Hurt |
| use ROCK instead of ROW-50 on common corruptions | ROW-50 · PartImageNet-C corruptions standard training | mCA: 35.8 -> 40.4 | Improved |
Closest papers by meaning
- Part-Based Models Improve Adversarial Robustness (ICLR 2023 · similarity 0.90)
- Harmonizing the object recognition strategies of deep neural networks with humans (NeurIPS 2022 · similarity 0.88)
- PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition (ECCV 2024 · similarity 0.87)
- Performance-optimized deep neural networks are evolving into worse models of inferotemporal visual cortex (NeurIPS 2023 · similarity 0.87)
- Generalisation in humans and deep neural networks (NeurIPS 2018 · similarity 0.87)
- Contour Integration Underlies Human-Like Vision (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.
