Research map / Adversarial robustness and security / Certified and empirical robustness
Trust Region Based Adversarial Attack on Neural Networks
Zhewei Yao, Amir Gholami, Peng Xu, Kurt Keutzer, Michael W. Mahoney
Check 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 |
|---|---|---|---|
| Apply trust-region attack instead of DeepFool | VGG16 · ImageNet, Linf best-class attack | relative perturbation, average / worst: 0.11% / 4.0% -> 0.10% / 1.4% | Improved |
| Apply trust-region attack instead of FGSM | VGG16 · ImageNet, Linf best-class attack | relative perturbation, average / worst: 0.18% / 5.1% -> 0.10% / 1.4% | Improved |
| Apply trust-region attack instead of DeepFool | WResNet · Cifar-10, Linf best-class attack | relative perturbation, average / worst: 0.66% / 3.34% -> 0.56% / 2.67% | Improved |
| Apply trust-region attack instead of FGSM | WResNet · Cifar-10, Linf best-class attack | relative perturbation, average / worst: 0.85% / 3.30% -> 0.56% / 2.67% | Improved |
| Apply trust-region attack instead of DeepFool | WResNet · Cifar-10, L2 best-class attack | relative perturbation, average / worst: 0.89% / 5.79% -> 0.73% / 4.43% | Improved |
| Replace fixed TR radius with adaptive radius (ablation) | TR attack on VGG16 · ImageNet, Linf best-class attack | worst relative perturbation: 1.4% -> 3.4% | Hurt |
| Replace fixed TR radius with adaptive radius (ablation) | TR attack on ResNet · Cifar-10, Linf best-class attack | average relative perturbation: 0.56% -> 0.85% | Hurt |
| Use second-order rather than first-order TR (ablation) | TR attack on MLP with Swish · Cifar-10, 10 iterations | accuracy after attack: 0.27 -> 0.20 | Improved |
| Use second-order rather than first-order TR (ablation) | TR attack on AlexNet with Swish · Cifar-10, 10 iterations | accuracy after attack: 0.46 -> 0.44 | Improved |
| Use second-order rather than first-order TR (ablation) | TR attack on AlexNet with Swish · Cifar-10, 2 iterations | accuracy after attack: 28.17 -> 26.97 | Improved |
| Use second-order rather than first-order TR (ablation) | TR attack on MLP with Swish · Cifar-10, 2 iterations | accuracy after attack: 33.7 -> 33.37 | Mixed |
Closest papers by meaning
- Adversarial Parameter Attack on Deep Neural Networks (ICML 2023 · similarity 0.86)
- Cross-Domain Transferability of Adversarial Perturbations (NeurIPS 2019 · similarity 0.86)
- Regional Adversarial Training for Better Robust Generalization (IJCV 2024 · similarity 0.86)
- Towards Transferable Adversarial Attacks with Centralized Perturbation (AAAI 2024 · similarity 0.85)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.85)
- AttackBench: Evaluating Gradient-based Attacks for Adversarial Examples (AAAI 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.
