Research map / Adversarial robustness and security / Certified and empirical robustness
Connecting the Digital and Physical World: Improving the Robustness of Adversarial Attacks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 cGAN-simulated physical image as base for adversarial noise (pix2pix) in addition to EOT | EOT attack · Physical-domain frontal view, 102 ImageNet images | target top-1: 0.78 -> 0.91 | Helped |
| use cycleGAN rather than pix2pix for the transformation (ablation) | D2P · Physical-domain frontal view, 102 ImageNet images | target top-1: 0.91 -> 0.85 | Hurt |
| use real physical image as base instead of simulated one (ablation) | D2P · Physical-domain frontal view, 102 ImageNet images | target top-1: 0.91 -> 0.84 | No clear effect |
| add EOT geometric transformations to BIM | BIM · Physical-domain frontal view, 102 ImageNet images | target top-1: 0.00 -> 0.78 | Helped |
| D2P vs EOT at 45 degree viewing angle | D2P-pix2pix · Viewing angle 45 deg | target top-1: 0.33 -> 0.62 | Helped |
| swap printer to high-DPI Xerox | D2P-pix2pix · Diff. printer | target top-1: 0.91 -> 0.97 | Helped |
| swap phone camera to Android | D2P-pix2pix · Diff. phone | target top-1: 0.91 -> 0.90 | No clear effect |
| transfer to other classifiers (ResNet) | D2P-pix2pix · Transfer to ResNet | target top-1: 0.91 -> 0.35 | Hurt |
| D2P vs EOT on transfer to ResNet | D2P-pix2pix · Transfer to ResNet | target top-1: 0.17 -> 0.35 | Helped |
| increase pix2pix training images from 50 to 300 (ablation) | pix2pix · Table 2 | SSIM: 0.37 -> 0.46 | Helped |
| increase BIM noise level eps 30 to 70 | BIM · Physical-domain frontal view, 102 ImageNet images | target top-1 (physical): 0.00 -> 0.01 | No clear effect |
Closest papers by meaning
- Robust Physical-World Attacks on Deep Learning Visual Classification (CVPR 2018 · similarity 0.91)
- Breaking Barriers in Physical-World Adversarial Examples: Improving Robustness and Transferability via Robust Feature (AAAI 2025 · similarity 0.91)
- Synthesizing Robust Adversarial Examples (ICML 2018 · similarity 0.91)
- Evading Defenses to Transferable Adversarial Examples by Translation-Invariant Attacks (CVPR 2019 · similarity 0.89)
- Improving the Transferability of Adversarial Samples With Adversarial Transformations (CVPR 2021 · similarity 0.89)
- On Adversarial Training without Perturbing all Examples (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.