Research map / Adversarial robustness and security / Black-box and transfer attacks
Feature Importance-Aware Transferable Adversarial Attacks
Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, Kui Ren
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 aggregate gradient (averaged over random-masked images) to the feature-level loss (ablation) | FIA feature-level attack · Normal models, Inc-v3 source | attack success rate | Improved |
| use aggregate gradient instead of raw clean-image gradient (ablation) | FIA · Normal models, Inc-v3 source | attack success rate | Improved |
| add feature importance weighting to unconstrained feature distortion (L1) (ablation) | feature-level attack · ImageNet-compatible | attack success rate | Improved |
| combine FIA with PIDIM | FIA · Normal models | attack success rate: 1% ~ 10% increase of success rate | Improved |
| replace FDA/NRDM with FIA as feature-level attack | NRDM, FDA · Normal models | attack success rate: 60.3% -> 83.5% | Improved |
| increase drop probability p_d to 0.5 (ablation) | FIA · Inc-v3 source, varied targets | attack success rate | Hurt |
| use p_d between 0.2 and 0.3 for normal targets, around 0.1 for defense targets (ablation) | FIA · Inc-v3 source | attack success rate | Improved |
| increase ensemble number N from 5 to 40 (ablation) | FIA aggregate gradient · Inc-v3 source | attack success rate | Improved |
| attack middle layers instead of early or late layers (ablation) | FIA · Inc-v3 and Vgg-16 sources | attack success rate | Improved |
| generate adversarial examples on an ensemble of models rather than single model | all attacks · Defense models | attack success rate: 68.0% -> 90.9% | Improved |
| add PITIDIM to FIA | FIA · Defense models, Vgg-16 source | attack success rate: 87.8% -> 74.7% | Hurt |
| add PITIDIM to FIA | FIA · Defense models, Inc-v3 source | attack success rate: 54.5% -> 64.8% | Improved |
Closest papers by meaning
- Boosting the Transferability of Adversarial Samples via Attention (CVPR 2020 · similarity 0.91)
- Boosting Adversarial Transferability through Augmentation in Hypothesis Space (CVPR 2025 · similarity 0.90)
- Rethinking Model Ensemble in Transfer-based Adversarial Attacks (ICLR 2024 · similarity 0.90)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.89)
- Pixel2Feature Attack (P2FA): Rethinking the Perturbed Space to Enhance Adversarial Transferability (ICML 2025 · similarity 0.89)
- Transferability Bound Theory: Exploring Relationship between Adversarial Transferability and Flatness (NeurIPS 2024 · 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.
