Research map / Adversarial robustness and security / Certified and empirical robustness
Removing Adversarial Noise in Class Activation Feature Space
Dawei Zhou, Nannan Wang, Chunlei Peng, Xinbo Gao, Xiaoyu Wang, Jun Yu, Tongliang Liu
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 class activation feature denoiser CAFD against AA-N | VGG-19 defense · CIFAR-10 AA-N | error rate: 30.61 -> 11.80 | Improved |
| add CAFD against BPDA adaptive attack, versus APE and HGD | VGG-19 defense · BPDA PGD-N | error rate: reduced by 49.85% and 39.83% on average | Improved |
| CAFD under white-box adaptive attack (defense leaked) | VGG-19 defense · White-box DDN-N | error rate: 97.85 -> 93.23 | No clear effect |
| CAFD' as local defense under gray-box PGD-N | VGG-19 defense · Gray-box PGD-N | error rate: 65.03 -> 56.76 | Improved |
| CAFD under gray-box DDN-N | VGG-19 defense · Gray-box DDN-N | error rate: 10.93 -> 10.90 | No clear effect |
| transfer CAFD to ResNet-50 target | ResNet-50 · SVHN PGD-N | error rate: 100 -> 11.01 | Improved |
| transfer CAFD to Wide-ResNet target | Wide-ResNet · SVHN PGD-N | error rate: 97.71 -> 21.33 | Improved |
| add class activation feature loss L_caf (ablation) | CAFD · CIFAR-10 attacks | accuracy | Improved |
| add adversarial (discriminator) loss L_adv (ablation) | CAFD · CIFAR-10 attacks | accuracy | Improved |
| use class activation feature loss instead of pixel-wise MSE loss (ablation) | CAFD · CIFAR-10 attacks | accuracy | Improved |
| use default CAFD loss form rather than variant CAFD' (ablation) | CAFD · SVHN PGD-N | error rate: 15.91 -> 14.64 | Improved |
| increase perturbation budget beyond 12/255 (ablation) | CAFD · CIFAR-10 PGD/AA | accuracy | Hurt |
Closest papers by meaning
- Feature Denoising for Improving Adversarial Robustness (CVPR 2019 · similarity 0.88)
- First Line of Defense: A Robust First Layer Mitigates Adversarial Attacks (AAAI 2025 · similarity 0.87)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.86)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.85)
- Identifying and Understanding Cross-Class Features in Adversarial Training (ICML 2025 · similarity 0.85)
- Theoretical Understanding of Learning from Adversarial Perturbations (ICLR 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.
