Research map / Adversarial robustness and security / Black-box and transfer attacks
CAG: A Real-Time Low-Cost Enhanced-Robustness High-Transferability Content-Aware Adversarial Attack Generator
Huy Phan, Yi Xie, Siyu Liao, Jie Chen, Bo Yuan
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 |
|---|---|---|---|
| Replace iterative optimization with a U-Net generator | CAG on CIFAR-10 · CIFAR-10 | Generation time: 12m56s -> 1.44s | Improved |
| Add perturbation dropout, increasing p from 0.0 to 0.3 (ablation) | CAG · CIFAR-10 black-box | ASR: 83.78% -> 89.24% | Improved |
| Add perturbation dropout, increasing p from 0.0 to 0.1 (ablation) | CAG · CIFAR-10 black-box | ASR: 83.78% -> 85.92% | Improved |
| Add perturbation dropout, increasing p from 0.1 to 0.2 (ablation) | CAG · CIFAR-10 black-box | ASR: 85.92% -> 87.85% | Improved |
| Add perturbation dropout, increasing p from 0.2 to 0.3 (ablation) | CAG · CIFAR-10 black-box | ASR: 87.85% -> 89.24% | Improved |
| Train one generator with label embeddings rather than separate generators for each target class | CAG · ImageNet 1000 targeted classes | Storage: 30 MB × 1000 -> 222 MB | Improved |
| Use a content-aware generator with CAM supervision instead of GAP U-Net | CAG · ImageNet 5 targeted classes, black-box | ASR: 85.66% -> 94.62% | Improved |
| Use a content-aware generator with CAM supervision instead of GAP U-Net | CAG · ImageNet 5 targeted classes, white-box | ASR: 98.09% -> 97.71% | Mixed |
| Constrain perturbations using CAM information | CAG · ImageNet, Pixel Deflection | Classification accuracy after defense: 27.63% | Improved |
| Use ensemble classifiers for ImageNet CAG training | CAG · ImageNet | Overfitting | Improved |
Closest papers by meaning
- NAG: Network for Adversary Generation (CVPR 2018 · similarity 0.86)
- Transferable Adversarial Attack based on Integrated Gradients (ICLR 2022 · similarity 0.86)
- AIM: Additional Image Guided Generation of Transferable Adversarial Attacks (AAAI 2025 · similarity 0.86)
- Dual-Flow: Transferable Multi-Target, Instance-Agnostic Attacks via $\textit{In-the-wild}$ Cascading Flow Optimization (NeurIPS 2025 · similarity 0.85)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.85)
- Improved Generation of Adversarial Examples Against Safety-aligned LLMs (NeurIPS 2024 · 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.
