Research map / Adversarial robustness and security / Black-box and transfer attacks
Transferable Adversarial Attack based on Integrated Gradients
Yi Huang, Adams Wai-Kin Kong
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 random piecewise linear path integrated gradients instead of straight-line integrated gradients (ablation) | TAIG · ImageNet val, untargeted, ε=0.03 | average attack success rate (%): 31.72 -> 52.82 | Improved |
| Use random piecewise linear path integrated gradients instead of straight-line integrated gradients (ablation) | TAIG · ImageNet val, untargeted, ε=0.05 | average attack success rate (%): 52.13 -> 79.83 | Improved |
| Use random piecewise linear path integrated gradients instead of straight-line integrated gradients (ablation) | TAIG · ImageNet val, untargeted, ε=0.1 | average attack success rate (%): 77.15 -> 96.90 | Improved |
| Use random piecewise linear path integrated gradients instead of straight-line integrated gradients (ablation) | TAIG · Advanced defenses, ε=16/255 | average attack success rate (%): 34.84 -> 70.82 | Improved |
| Use random piecewise linear path integrated gradients instead of straight-line integrated gradients (ablation) | TAIG · ImageNet val, targeted, ε=0.10 | average attack success rate (%): 5.15 -> 26.03 | Improved |
| Replace standard IFGSM gradients with integrated-gradient updates | ILA · ImageNet val, untargeted, ε=0.1 | average attack success rate (%): 91.36 -> 96.57 | Improved |
| Use TAIG-R rather than IFGSM as the backend | LinBP · ImageNet val, untargeted, ε=0.1 | average attack success rate (%): 94.50 -> 99.60 | Improved |
| Use TAIG-R rather than MIFGSM gradients | DI · ImageNet val, untargeted, ε=0.1 | average attack success rate (%): 70.16 -> 98.46 | Improved |
| Combine TAIG-R with other methods in targeted attacks | TAIG-R · Targeted attacks | transferability | Hurt |
| Combine TAIG-S with ILA in targeted attacks | TAIG-S · Targeted attacks | transferability | Hurt |
| Increase the number of AOA iterations from 100 to 300 (ablation) | AOA · ImageNet val, ε=16/255 | average attack success rate (%): 32.07 to 32.19 | No clear effect |
| Increase the number of gradient calculations (ablation) | SI and LinBP · ImageNet val, untargeted attacks | average attack success rate (%) | Improved |
Closest papers by meaning
- Improving Integrated Gradient-based Transferable Adversarial Examples by Refining the Integration Path (AAAI 2025 · similarity 0.93)
- Transferable Adversarial Perturbations (ECCV 2018 · similarity 0.90)
- Evading Defenses to Transferable Adversarial Examples by Translation-Invariant Attacks (CVPR 2019 · similarity 0.90)
- Yet Another Intermediate-Level Attack (ECCV 2020 · similarity 0.89)
- Enhancing Adversarial Example Transferability With an Intermediate Level Attack (ICCV 2019 · similarity 0.89)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 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.
