Research map / Adversarial robustness and security / Black-box and transfer attacks
Transferable Adversarial Perturbations
Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, Yong Yang
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 TAP instead of BIM to craft perturbations on Inception-V3 | black-box attack on Inception-V4 · ImageNet-compatible, source Inception-V3 | recognition accuracy: 67.7% -> 22.1% | Improved |
| use TAP instead of FGSM on Inception-V3 | black-box attack on Inception-ResNet-V2 · ImageNet-compatible, source Inception-V3 | recognition accuracy: 72.9% -> 24.7% | Improved |
| use TAP instead of BIM on Inception-V3 | adversarially trained adv-v3 · ImageNet-compatible, source Inception-V3 | recognition accuracy: 82.4% -> 52.8% | Improved |
| add the intermediate feature-distance term (lambda 0.05, eta 0) (ablation) | BIM · Inception-V3 source, standard and ensemble models | recognition accuracy | Improved |
| add the feature-distance term (lambda 0.05, eta 0) (ablation) | BIM · adversarially trained models | recognition accuracy | Improved |
| add the smooth regularization term (lambda 0, eta 10^3) (ablation) | BIM · adversarially trained models | recognition accuracy | Improved |
| add the smooth regularization term (lambda 0, eta 10^3) (ablation) | BIM · Inception-V3 white-box | recognition accuracy | Hurt |
| add the smooth regularization term (lambda 0, eta 10^3) (ablation) | BIM · Inception-V4 and Inception-ResNet-V2 | recognition accuracy | Hurt |
| increase lambda (ablation) | TAP · adversarially trained vs standard models | recognition accuracy | Mixed |
| increase eta (ablation) | TAP · adversarially trained vs standard models | recognition accuracy | Mixed |
| set power normalization alpha to 0.5 instead of other values (ablation) | TAP · defense models | recognition accuracy | Improved |
| increase smoothing kernel size s above 3 (ablation) | TAP · all models | recognition accuracy | Hurt |
Closest papers by meaning
- Enhancing Adversarial Example Transferability With an Intermediate Level Attack (ICCV 2019 · similarity 0.93)
- Perturbing Across the Feature Hierarchy to Improve Standard and Strict Blackbox Attack Transferability (NeurIPS 2020 · similarity 0.92)
- Transferable Perturbations of Deep Feature Distributions (ICLR 2020 · similarity 0.92)
- Evading Defenses to Transferable Adversarial Examples by Translation-Invariant Attacks (CVPR 2019 · similarity 0.92)
- Towards Transferable Adversarial Attacks with Centralized Perturbation (AAAI 2024 · similarity 0.90)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.90)
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.
