Research map / Adversarial robustness and security / Black-box and transfer attacks
Boosting Transferability of Targeted Adversarial Examples via Hierarchical Generative Networks
Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu
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 instance-specific Logit attack with the conditional generator | Logit · NeurIPS 1k, 8 targets, Inv3 source, target Inv4 | targeted success rate: 5.6 -> 66.9 | Improved |
| add smooth tanh projection and adaptive Gaussian smoothing over CD-AP | CD-AP · NeurIPS 1k, 8 targets, Inv3 source, target Inv3ens3 | targeted success rate: 1.5 -> 37.2 | Improved |
| use one conditional model instead of 8 separate CD-AP models | CD-AP · NeurIPS 1k, 8 targets, Inv3 source, target Inv4 | targeted success rate: 57.6 -> 66.9 | Improved |
| replace Gaussian smoothing of CD-AP-gs with the full method | CD-AP-gs · NeurIPS 1k, 8 targets, Inv3 source, target Inv4 | targeted success rate: 31.3 -> 66.9 | Improved |
| hierarchical k-DPP partition into 20 models of 50 classes | Conditional generator · NeurIPS 2017 competition, 1000 target classes, target Inv4 | targeted success rate: 4.7 -> 35.9 | Improved |
| reduce perturbation budget from eps 16 to eps 8 (ablation) | Ours · Targeted black-box, Inv3 source, target VGG-16 | targeted success rate: 61.9 -> 36.1 | Hurt |
| reduce eps from 16 to 8 on R152 target (ablation) | Ours · Targeted black-box, Inv3 source, target R152 | targeted success rate: 49.6 -> 16.3 | Hurt |
| compare Ours against Logit at eps 8 | Logit · Targeted black-box, Inv3 source, target VGG-16, eps=8 | targeted success rate: 2.1 -> 36.1 | Improved |
| train generator on MS-COCO instead of ImageNet (ablation) | Ours · Targeted black-box, Inv3 substitute, target DN | targeted success rate: 79.9 -> 70.3 | Hurt |
| train generator on Comics instead of ImageNet (ablation) | Ours · Targeted black-box, Inv3 substitute, target DN | targeted success rate: 79.9 -> 60.4 | Hurt |
| use diverse conditional classes from k-DPP instead of cosine-close classes (ablation) | Ours · ImageNet target class subsets | loss convergence and transferability | Improved |
| increase the number of target classes handled by a single generator (ablation) | Ours · ImageNet target classes | targeted success rate | Hurt |
Closest papers by meaning
- On Generating Transferable Targeted Perturbations (ICCV 2021 · similarity 0.93)
- Towards Transferable Targeted Attack (CVPR 2020 · similarity 0.93)
- AIM: Additional Image Guided Generation of Transferable Adversarial Attacks (AAAI 2025 · similarity 0.92)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.89)
- Everywhere Attack: Attacking Locally and Globally to Boost Targeted Transferability (AAAI 2025 · similarity 0.89)
- Boosting Adversarial Transferability through Augmentation in Hypothesis Space (CVPR 2025 · similarity 0.89)
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.
