Research map / Adversarial robustness and security / Certified and empirical robustness
AGKD-BML: Defense Against Adversarial Attack by Attention Guided Knowledge Distillation and Bi-Directional Metric Learning
Hong Wang, Yuefan Deng, Shinjae Yoo, Haibin Ling, Yuewei Lin
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 bidirectional metric learning in place of single-directional metric learning (ablation) | AGKD-BML · CIFAR-10 | PGD-20 accuracy: 51.59% -> 60.51% | Improved |
| add attention guided knowledge distillation alone to the adversarial training baseline (ablation) | adversarial training · CIFAR-10 | PGD-20 accuracy: 45.57% -> 65.93% | Improved |
| add AGKD to BML (ablation) | BML · CIFAR-10 | PGD-20 accuracy: 60.51% -> 71.02% | Improved |
| add BML to AGKD (ablation) | AGKD · CIFAR-10 | PGD-20 accuracy: 65.93% -> 71.02% | Improved |
| add BML to AGKD (ablation) | AGKD · CIFAR-10 | CW-20 accuracy: 60.71% -> 63.67% | Improved |
| add BML to AGKD (ablation) | AGKD · CIFAR-10 | FGSM accuracy: 75.57% -> 76.69% | Improved |
| train with 7-step attack instead of 2-step | AGKD-BML · CIFAR-10 | AutoAttack accuracy: 37.07% -> 50.59% | Improved |
| train with 7-step attack instead of 2-step | AGKD-BML · CIFAR-10 | PGD-20 accuracy: 71.02% -> 57.30% | Hurt |
| use class-irrelevant attention (AGKD) instead of class-relevant Grad-CAM attention of original class (ablation) | AGKD · CIFAR-10 | PGD-20 accuracy: 59.06% -> 65.93% | Improved |
| use CAM-avg class-irrelevant attention instead of AGKD attention (ablation) | AGKD · CIFAR-10 | PGD-20 accuracy: 65.93% -> 66.55% | No clear effect |
| train FS with a 2-step attack to match AGKD-BML training time | feature scattering · CIFAR-10 | PGD-20 accuracy: 70.51% vs 71.02% | No clear effect |
| use WRN-34-10 instead of WRN-28-10 with decay at 150 (ablation) | AGKD-BML-10 · CIFAR-10 | AutoAttack accuracy: 50.73% -> 51.63% | Improved |
Closest papers by meaning
- Revisiting Adversarial Training at Scale (CVPR 2024 · similarity 0.86)
- Impartial Adversarial Distillation: Addressing Biased Data-Free Knowledge Distillation via Adaptive Constrained Optimization (AAAI 2024 · similarity 0.86)
- Out of Thin Air: Exploring Data-Free Adversarial Robustness Distillation (AAAI 2024 · similarity 0.86)
- Robust Distillation via Untargeted and Targeted Intermediate Adversarial Samples (CVPR 2024 · similarity 0.85)
- KDAT: Inherent Adversarial Robustness via Knowledge Distillation with Adversarial Tuning for Object Detection Models (AAAI 2025 · similarity 0.85)
- AGS: Affordable and Generalizable Substitute Training for Transferable Adversarial Attack (AAAI 2024 · 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.
