Research map / Adversarial robustness and security / Black-box and transfer attacks
Data-Free Universal Adversarial Perturbation and Black-Box Attack
Chaoning Zhang, Philipp Benz, Adil Karjauv, 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 CosSim minimization loss instead of CosSim-max (ablation) | Cosine-UAP · ImageNet val (ResNet152) | fooling ratio: 80.2 -> 90.2 | Improved |
| use CosSim loss instead of CosSim-max (ablation) | Cosine-UAP · ImageNet val (VGG19) | fooling ratio: 91.1 -> 96.4 | Improved |
| use jigsaw images with variable frequency instead of fixed frequency (ablation) | Data-free Cosine-UAP · ImageNet val (avg 5 models) | fooling ratio: 79.4 -> 84.08 | Improved |
| use jigsaw images instead of Gaussian noise (ablation) | Data-free Cosine-UAP · ImageNet val (avg 5 models) | fooling ratio: 67.9 -> 84.08 | Improved |
| use jigsaw images instead of flat images (ablation) | Data-free Cosine-UAP · ImageNet val (avg 5 models) | fooling ratio: 62.1 -> 84.08 | Improved |
| randomly resize the UAP during training (ablation) | Cosine-UAP · ImageNet val (224 input) | fooling ratio: 89.48 -> 86.16 | Hurt |
| randomly resize the UAP during training (ablation) | Cosine-UAP · ImageNet val (112 input) | fooling ratio: 58.57 -> 62.05 | Improved |
| randomly resize the UAP during training (transfer VGG16 to ResNet152) (ablation) | Cosine-UAP · ImageNet val | fooling ratio: 38.37 -> 44.27 | Improved |
| remove high-frequency content with Fourier transform | checkerboard attack · ImageNet black-box | accuracy under attack: 25.3% -> 20.8% | Improved |
| use checkerboard instead of uniform noise pattern (no HF removal) | optimization-free attack · ImageNet black-box | accuracy under attack: 67.3% -> 25.3% | Improved |
| use optimization-free checkerboard instead of Prototypical* | data-free black-box attack · ImageNet black-box | accuracy under attack: 31.2% -> 25.3% | Improved |
| reduce training samples from full set to 32 with augmentation | Cosine-UAP · ImageNet val (VGG16) | fooling ratio: 97.4 -> 93.5 | Hurt |
Closest papers by meaning
- Data-free Universal Adversarial Perturbation with Pseudo-semantic Prior (CVPR 2025 · similarity 0.92)
- Data-Free Universal Attack by Exploiting the Intrinsic Vulnerability of Deep Models (AAAI 2025 · similarity 0.92)
- Generate Universal Adversarial Perturbations for Few-Shot Learning (NeurIPS 2024 · similarity 0.89)
- Improving Generalization of Universal Adversarial Perturbation via Dynamic Maximin Optimization (AAAI 2025 · similarity 0.89)
- Democratic Training Against Universal Adversarial Perturbations (ICLR 2025 · similarity 0.87)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · 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.
