Research map / Adversarial robustness and security / Black-box and transfer attacks
Patch-wise Attack for Fooling Deep Neural Network
Lianli Gao, Qilong Zhang, Jingkuan Song, Xianglong Liu, Heng Tao Shen
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 patch-wise project kernel on top of MI-FGSM-style baseline (PI-FGSM vs MI-FGSM) | MI-FGSM, source Inc-v3 · NT target Dense-161 | success rate (%): 45.8 -> 61.7 | Improved |
| add PI to DMI-FGSM (DMPI-FGSM) | DMI-FGSM, source Inc-v3 · NT target Inc-v4 | success rate (%): 78.9 -> 81.8 | Improved |
| add PI to DMI-FGSM (DMPI-FGSM) | DMI-FGSM, source IncRes-v2 · NT target Inc-v3 | success rate (%): 86.8 -> 92.2 | Improved |
| add PI to DTMI-FGSM (DTPI-FGSM) | DTMI-FGSM, ensemble source · EAT target Inc-v3_ens3 | success rate (%): 81.2 -> 89.3 | Improved |
| add PI to DTMI-FGSM (DTPI-FGSM) | DTMI-FGSM, ensemble source · FD target ResNeXt_DA | success rate (%): 6.1 -> 11.7 | Improved |
| add PI to DTMI-FGSM when single source IncRes-v2 | DTMI-FGSM · EAT target IncRes-v2_ens | success rate (%): 62.9 -> 62.3 | No clear effect |
| add PI to MI-FGSM with source IncRes-v2 (PI-FGSM vs MI-FGSM) | MI-FGSM · NT target Inc-v3 | success rate (%): 77.6 -> 76.1 | Hurt |
| increase project factor gamma from 1.0 to 16.0 (ablation) | PI-FGSM, source Inc-v3 · avg of 8 black-box models | average success rate (%): 24.5 -> 46.2 | Improved |
| increase project factor gamma from 3.0 to 16.0 (ablation) | DMPI-FGSM, source Inc-v3 · NT average | average success rate (%): 72.6 -> 57.4 | Hurt |
| increase project factor gamma from 1.0 to 16.0 (ablation) | PI-FGSM, source Res152_B · FD average | average success rate (%): 70.7 -> 58.5 | Hurt |
| increase iterations T from 5 to 30 (ablation) | PI-FGSM, source Inc-v3 · NT average | average success rate (%): 55.2 -> 53.8 | No clear effect |
| increase iterations T from 5 to 30 (ablation) | PI-FGSM, source Inc-v3 · EAT average | average success rate (%): 23.0 -> 28.3 | Improved |
Closest papers by meaning
- Decision-based Black-box Attack Against Vision Transformers via Patch-wise Adversarial Removal (NeurIPS 2022 · similarity 0.88)
- Sparse and Imperceivable Adversarial Attacks (ICCV 2019 · similarity 0.87)
- Beyond ImageNet Attack: Towards Crafting Adversarial Examples for Black-box Domains (ICLR 2022 · similarity 0.87)
- SlowFormer: Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers (CVPR 2024 · similarity 0.86)
- PAD: Patch-Agnostic Defense against Adversarial Patch Attacks (CVPR 2024 · similarity 0.85)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.85)
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.
