Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Zero-Shot Attribute Attacks on Fine-Grained Recognition Models
Nasim Shafiee, Ehsan Elhamifar
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 regularization loss Lreg to the ranking loss (ablation) | CAUP on DAZLE · CUB (l2-norm 6) | seen/unseen fooling fraction: 0.24/0.18 -> 0.53/0.40 | Improved |
| add utility loss Lutil to the ranking loss (ablation) | CAUP on DAZLE · CUB (l2-norm 6) | seen/unseen fooling fraction: 0.24/0.18 -> 0.90/0.82 | Improved |
| add both Lreg and Lutil to the ranking loss (ablation) | CAUP on DAZLE · CUB (l2-norm 6) | seen/unseen fooling fraction: 0.24/0.18 -> 0.96/0.89 | Improved |
| add both Lreg and Lutil to the ranking loss (ablation) | CAUP on DAZLE · AWA2 (l2-norm 6) | seen/unseen fooling fraction: 0.05/0.11 -> 0.76/0.73 | Improved |
| add both Lreg and Lutil to the ranking loss (ablation) | CAUP on DAZLE · SUN (l2-norm 6) | seen/unseen fooling fraction: 0.38/0.22 -> 0.75/0.55 | Improved |
| add Lutil to the ranking loss (ablation) | CAUP on DAZLE · SUN (l2-norm 6) | seen/unseen fooling fraction: 0.38/0.22 -> 0.30/0.26 | Mixed |
| compose AUPs with class-attribute weights instead of uniform weights (ablation) | CAUP on DAZLE · CUB, AWA2, SUN | fooling percentage vs l2-norm | Improved |
| use compositional attribute-based perturbations instead of single universal perturbation (UAP, GAP) | CAUP on DAZLE · CUB (l2-norm 6) | seen fooling percentage: 35 -> 96 | Improved |
| use compositional attribute-based perturbations instead of UAP, GAP | CAUP on DAZLE · AWA2 (l2-norm 6) | seen fooling percentage: 11 -> 89 | Improved |
| use CAUP instead of UAP on a robust model | CAUP on DCN · AWA2 (l2-norm 6) | seen fooling percentage: 02 -> 08 | Improved |
| use CAUP instead of UAP | CAUP on CNZSL · SUN (l2-norm 6) | seen fooling percentage: 29 -> 22 | Hurt |
| use CAUP under l-infinity norm 0.06 instead of GAP | CAUP · 12 model-dataset-pair cases | fooling percentage | Mixed |
Closest papers by meaning
- Fine-Grained Generalized Zero-Shot Learning via Dense Attribute-Based Attention (CVPR 2020 · similarity 0.88)
- Compositional Zero-Shot Learning via Fine-Grained Dense Feature Composition (NeurIPS 2020 · similarity 0.86)
- Adversarial Fine-Grained Composition Learning for Unseen Attribute-Object Recognition (ICCV 2019 · similarity 0.85)
- Generalizable Data-Free Objective for Crafting Universal Adversarial Perturbations (TPAMI 2018 · similarity 0.85)
- Generate Universal Adversarial Perturbations for Few-Shot Learning (NeurIPS 2024 · similarity 0.85)
- Data-Free Universal Attack by Exploiting the Intrinsic Vulnerability of Deep Models (AAAI 2025 · similarity 0.84)
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.
