Research map / Adversarial robustness and security / Black-box and transfer attacks
Training Meta-Surrogate Model for Transferable Adversarial Attack
Yunxiao Qin, Yuanhao Xiong, Jinfeng Yi, Cho-Jui Hsieh
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 |
|---|---|---|---|
| train and attack a meta-surrogate rather than attack the original surrogate directly | CIFAR-10 transfer attack · CIFAR-10, MobileNet-V3 | attack success rate: 90.8% -> 91.8% | Improved |
| train and attack a meta-surrogate rather than attack the original surrogate directly | CIFAR-10 transfer attack · CIFAR-10, adversarially trained ResNet-18 | attack success rate: 96.1% -> 98.4% | Improved |
| train and attack a meta-surrogate rather than attack the original surrogate directly | ImageNet transfer attack · ImageNet, Inc-V3 ens3 | attack success rate: 71.1% -> 75.5% | Improved |
| add the arctan gradient component to Customized PGD (ablation) | MTA-PGD · CIFAR-10, MobileNet-V3 | attack success rate: 70.0% -> 91.8% | Improved |
| add the sign gradient component to Customized PGD (ablation) | MTA-PGD · CIFAR-10, MobileNet-V3 | attack success rate: 90.0% -> 91.8% | Improved |
| replace ResNet-13 with DenseNet-22BC as the MSM backbone (ablation) | MTA-PGD · CIFAR-10, MobileNet-V3 | attack success rate: 91.8% -> 86.9% | Hurt |
| use PGD-L2 instead of Customized PGD during MSM training (ablation) | MTA-PGD · CIFAR-10, MobileNet-V3 | attack success rate: 80.8% -> 91.8% | Hurt |
| use PGD-L1 instead of Customized PGD during MSM training (ablation) | MTA-PGD · CIFAR-10, MobileNet-V3 | attack success rate: 81.5% -> 91.8% | Hurt |
| increase γ1 from 0.01 to 0.05 (ablation) | Customized PGD training for MTA-PGD · CIFAR-10, MobileNet-V3 | attack success rate: 90.5% vs 91.8% | Hurt |
| increase γ2 from 0.01 to 0.05 (ablation) | Customized PGD training for MTA-PGD · CIFAR-10, MobileNet-V3 | attack success rate: 86.7% vs 91.8% | Hurt |
| increase the number of training attack steps from below 7 toward 7 (ablation) | MTA-PGD · CIFAR-10 target models | attack success rate | Improved |
| increase the number of training attack steps beyond 7 (ablation) | MTA-PGD · CIFAR-10 target models | attack success rate | Hurt |
Closest papers by meaning
- Seeking Flat Minima Over Diverse Surrogates for Improved Adversarial Transferability: A Theoretical Framework and Algorithmic Instantiation (TPAMI 2026 · similarity 0.93)
- Generalizable Black-Box Adversarial Attack With Meta Learning (TPAMI 2023 · similarity 0.92)
- LRS: Enhancing Adversarial Transferability through Lipschitz Regularized Surrogate (AAAI 2024 · similarity 0.91)
- Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training (CVPR 2025 · similarity 0.89)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.89)
- PubDef: Defending Against Transfer Attacks From Public Models (ICLR 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.
