Research map / Adversarial robustness and security / Certified and empirical robustness
MixACM: Mixup-Based Robustness Transfer via Distillation of Activated Channel Maps
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 mixup to ACM+soft-label distillation (ablation) | MixACM · CIFAR-10 WRN-34-10 | PGD robust accuracy: 49.53 -> 56.65 | Helped |
| add mixup to ACM+soft-label distillation (ablation) | MixACM · CIFAR-10 WRN-34-10 | clean accuracy: 91.17 -> 90.76 | Hurt |
| add teacher soft labels (KD) to ACM loss (ablation) | MixACM · CIFAR-10 WRN-34-10 | PGD robust accuracy: 48.38 -> 49.53 | Helped |
| add mixup to ACM loss + mixup training (ablation) | MixACM w/o KD · CIFAR-10 WRN-34-10 | PGD robust accuracy: 48.38 -> 52.29 | Helped |
| add ACM loss to soft-label KD (ablation) | KD only · CIFAR-10 WRN-34-10 | PGD robust accuracy: 0.21 -> 48.38 | Helped |
| distill mid-level features in addition to high-level only (ablation) | MixACM · CIFAR-10 | PGD robust accuracy: 33.13 -> 52.79 | Helped |
| use high-level features instead of mid-level only (ablation) | MixACM · CIFAR-10 | clean accuracy: 88.30 -> 91.18 | Helped |
| add first layer distillation (ablation) | MixACM · CIFAR-10 | PGD robust accuracy: 56.15 -> 56.65 | No clear effect |
| use ACM loss instead of direct feature distillation, different channels (ablation) | MixACM · CIFAR-10 WRN28-10 to WRN28-5 | PGD robust accuracy: 44.95 -> 48.75 | Helped |
| use ACM loss instead of direct feature distillation (ablation) | MixACM · CIFAR-100 transfer learning | PGD robust accuracy: 16.20 -> 24.14 | Helped |
| use adaptive max pool on teacher maps instead of affine mapping (ablation) | MixACM · CIFAR-10 WRN28-10 to WRN28-5 | PGD robust accuracy: 23.51 -> 48.75 | Helped |
| add random augmentation (ablation) | MixACM · ImageNet ResNet-50, eps 4/255 | PGD100 robust accuracy: 31.39 -> 33.63 | Helped |
Closest papers by meaning
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.86)
- Initialization Matters for Adversarial Transfer Learning (CVPR 2024 · similarity 0.86)
- Out of Thin Air: Exploring Data-Free Adversarial Robustness Distillation (AAAI 2024 · similarity 0.86)
- Improving Transferable Targeted Attacks with Feature Tuning Mixup (CVPR 2025 · similarity 0.86)
- MixAT: Combining Continuous and Discrete Adversarial Training for LLMs (NeurIPS 2025 · similarity 0.86)
- Adversarially Robust Hypothesis Transfer Learning (ICML 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.