Research map / Adversarial robustness and security / Attacks on text and graphs
CD-UAP: Class Discriminative Universal Adversarial Perturbation
Chaoning Zhang, Philipp Benz, Tooba Imtiaz, In-So Kweon
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 |
|---|---|---|---|
| use CD-UAP with bounded-logit loss for all-class attack | ImageNet AC-UAP · ImageNet | fooling rate: 93.7 vs 83.7 (VGG16); 94.2 vs 80.1 (VGG19); 90.2 vs 84.0 (ResNet152); 85.9 vs 82.7 (Inception-V3) | Improved |
| use half-half targeted/non-targeted batch sampling (ablation) | CD-UAP · CIFAR100, classes 0–4 | AAD_t: 8.00 -> 48.80 | Improved |
| use half-half targeted/non-targeted batch sampling (ablation) | CD-UAP · CIFAR100, classes 0–4 | AAD_nt: 1.79 -> 17.31 | Hurt |
| remove half-half sampling and compensate with α=19, β=1 (ablation) | CD-UAP · CIFAR100, classes 0–4 | AAD_t: 48.80 -> 48.20 | Hurt |
| remove half-half sampling and compensate with α=19, β=1 (ablation) | CD-UAP · CIFAR100, classes 0–4 | AAD_nt: 17.31 -> 19.43 | Hurt |
| train using only targeted-class samples rather than both groups (ablation) | CD-UAP · CIFAR100, classes 0–4 | AAD_nt: 17.31 -> 53.67 | Hurt |
| use bounded-logit loss rather than unbounded-logit loss for targeted samples (ablation) | CD-UAP · CIFAR100 | AAD_nt: 46.02 -> 7.16 | Improved |
| use cross-entropy loss for non-targeted samples instead of bounded-logit loss (ablation) | CD-UAP with targeted bounded-logit loss · CIFAR100 | AAD_nt: 17.79 -> 17.31 | Improved |
| target two superclasses simultaneously instead of one (ablation) | CD-UAP · CIFAR100 superclasses | Δ_AAD | Hurt |
| transfer a perturbation between networks in the same architecture family | CD-UAP · ImageNet transferability | Δ_AAD: 51.65* vs 25.92 (VGG16 to VGG19); 54.81* vs 30.92 (VGG19 to VGG16) | Mixed |
Closest papers by meaning
- Learning Universal Adversarial Perturbation by Adversarial Example (AAAI 2022 · similarity 0.92)
- Data-free Universal Adversarial Perturbation with Pseudo-semantic Prior (CVPR 2025 · similarity 0.90)
- Improving Generalization of Universal Adversarial Perturbation via Dynamic Maximin Optimization (AAAI 2025 · similarity 0.89)
- Data-Free Universal Attack by Exploiting the Intrinsic Vulnerability of Deep Models (AAAI 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 0.87)
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.
