Research map / Adversarial robustness and security / Black-box and transfer attacks
Defending Against Universal Adversarial Patches by Clipping Feature Norms
Cheng Yu, Jiansheng Chen, Youze Xue, Yuyang Liu, Weitao Wan, Jiayu Bao, Huimin Ma
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 FNC layers (alpha=1.0) | ResNet-50 · ImageNet AdvP | adversarial accuracy: 0.3% -> 58.6% | Improved |
| add FNC layers (alpha=1.0) | ResNet-50 · ImageNet clean | clean accuracy: 76.1% -> 72.4% | Hurt |
| add FNC layers (alpha=1.0) | MobileNet-V2 · ImageNet LaVAN | adversarial accuracy: 4.8% -> 52.0% | Improved |
| add FNC on top of adversarial training DOA | DOA · CIFAR10 AdvP | adversarial accuracy: 46.9% -> 68.0% | Improved |
| increase clipping alpha from 1.0 to 1.1 (ablation) | FNC · ImageNet AdvP | adversarial accuracy: 58.6% -> 59.6% (ResNet-50); 51.9% -> 49.5% (MobileNet-V2) | Mixed |
| increase alpha too large or use too few FNC layers (ablation) | FNC · ImageNet LaVAN (ResNet-50) | adversarial accuracy | Hurt |
| use patch with aspect ratio 1:4 instead of 1:1 (ablation) | FNC · ImageNet LaVAN (ResNet-50) | adversarial accuracy: 58.3% -> 55.4% | Hurt |
| use single-image location-independent patch instead of universal patch (ablation) | FNC · ImageNet (ResNet-50) | adversarial accuracy: 58.3% -> 50.6% | Hurt |
| change attack target class from toaster to water jug (ablation) | FNC · ImageNet LaVAN (ResNet-50) | adversarial accuracy: 58.3% -> 63.3% | No clear effect |
| use FNC instead of PatchGuard on large-scale data | FNC · ImageNet AdvP (ResNet-50) | adversarial accuracy: 33.8% -> 58.6% | Improved |
Closest papers by meaning
- Defending Against Universal Attacks Through Selective Feature Regeneration (CVPR 2020 · similarity 0.87)
- Large Norms of CNN Layers Do Not Hurt Adversarial Robustness (AAAI 2021 · similarity 0.86)
- A Unified, Resilient, and Explainable Adversarial Patch Detector (CVPR 2025 · similarity 0.86)
- Breaking Barriers in Physical-World Adversarial Examples: Improving Robustness and Transferability via Robust Feature (AAAI 2025 · similarity 0.85)
- Data-Free Universal Attack by Exploiting the Intrinsic Vulnerability of Deep Models (AAAI 2025 · similarity 0.85)
- Pandora's Box: Towards Building Universal Attackers against Real-World Large Vision-Language Models (NeurIPS 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.
