Research map / Adversarial robustness and security / Certified and empirical robustness
Universal Adversarial Perturbation via Prior Driven Uncertainty Approximation
Hong Liu, Rongrong Ji, Jie Li, Baochang Zhang, Yue Gao, Yongjian Wu, Feiyue Huang
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 virtual epistemic uncertainty (MC dropout activation loss) (ablation) | GD-UAP · ImageNet val, VGG-F | fooling rate (%): 85.96 -> 87.75 | Improved |
| add virtual epistemic uncertainty (MC dropout activation loss) (ablation) | GD-UAP · ImageNet val, GoogleNet | fooling rate (%): 51.61 -> 61.41 | Improved |
| add texture-bias prior and style loss (ablation) | GD-UAP · ImageNet val, ResNet-50 | fooling rate (%): 35.59 -> 63.71 | Improved |
| add texture-bias prior and style loss (ablation) | GD-UAP · ImageNet val, VGG-16 | fooling rate (%): 45.47 -> 51.63 | Improved |
| add texture-bias prior on top of epistemic uncertainty (PD-UA vs UA) (ablation) | UA · ImageNet val, ResNet-50 | fooling rate (%): 38.87 -> 65.84 | Improved |
| add epistemic uncertainty to GD-UAP+P to get PD-UA (ablation) | GD-UAP+P · ImageNet val, GoogleNet | fooling rate (%): 58.37 -> 67.12 | Improved |
| use a texture image directly as the perturbation (no training) | PP · ImageNet val, ResNet-150 | fooling rate (%): 28.87 -> 24.12 | Mixed |
| use texture prior initialization instead of pre-trained-perturbation prior (ablation) | PD-UA · ImageNet val, GoogleNet (Table 4) | fooling rate (%): 60.58 -> 65.23 | Improved |
| use low-pass filter instead of high-pass filter in LPFM (ablation) | PD-UA · ImageNet val, ResNet-50 | fooling rate (%): 49.53 -> 65.84 | Improved |
| use LPFM instead of l2-norm gradient normalization (ablation) | PD-UA · ImageNet val, VGG-16 | fooling rate (%): 26.31 -> 53.09 | Improved |
| increase pyramid levels from 1 to 3 (ablation) | LPFM · ImageNet val, VGG-16 | fooling rate (%): 50.15 -> 53.09 | Improved |
| increase MC dropout samples T beyond 10 (ablation) | PD-UA · ImageNet val, ResNet-50 | fooling rate (%) | No clear effect |
Closest papers by meaning
- Improving Generalization of Universal Adversarial Perturbation via Dynamic Maximin Optimization (AAAI 2025 · similarity 0.89)
- Data-free Universal Adversarial Perturbation with Pseudo-semantic Prior (CVPR 2025 · similarity 0.89)
- Generate Universal Adversarial Perturbations for Few-Shot Learning (NeurIPS 2024 · similarity 0.88)
- Democratic Training Against Universal Adversarial Perturbations (ICLR 2025 · similarity 0.88)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · similarity 0.87)
- Data-Free Universal Attack by Exploiting the Intrinsic Vulnerability of Deep Models (AAAI 2025 · similarity 0.87)
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.
