Research map / Adversarial robustness and security / Attacks on text and graphs
SemanticAdv: Generating Adversarial Examples via Attribute-conditioned Image Editing
Haonan Qiu, Chaowei Xiao, Lei Yang, Xinchen Yan, Honglak Lee, Bo Li
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 feature-space interpolation (at f0) instead of attribute-space interpolation (ablation) | SemanticAdv · Face verification R-101-S, G-FPR=T-FPR=1e-3 | attack success rate (%): 0.08 -> 100.00 | Improved |
| interpolate on f0 rather than f-2 (ablation) | SemanticAdv · R-101-S, G-FPR=T-FPR=1e-4 | attack success rate (%): 59.53 -> 99.45 | Improved |
| interpolate on f0 rather than f2 (ablation) | SemanticAdv · R-101-S, G-FPR=T-FPR=1e-4 | attack success rate (%): 73.52 -> 99.45 | Improved |
| interpolate on f0 rather than f1 (ablation) | SemanticAdv · R-101-S, G-FPR=T-FPR=1e-4 | attack success rate (%): 97.58 -> 99.45 | Improved |
| interpolate on f0 rather than f-1 (ablation) | SemanticAdv · R-101-S, G-FPR=T-FPR=1e-3 | attack success rate (%): 100.00 -> 100.00 | No clear effect |
| use SemanticAdv instead of CW as the attacker | Face++ black-box API · Face++ (T-FPR=1e-3) | attack success rate (%): 37.24 -> 67.69 | Improved |
| use SemanticAdv instead of M-DI2-FGSM | Face++ black-box API · Face++ (T-FPR=1e-4) | attack success rate (%): 33.67 -> 48.21 | Improved |
| use SemanticAdv instead of M-DI2-FGSM | AliYun black-box API · AliYun (T-FPR=1e-4) | attack success rate (%): 18.00 -> 19.50 | Improved |
| add attribute-conditioned editing before CW (StarGAN+CW) | CW · AliYun (T-FPR=1e-4) | attack success rate (%): 9.50 -> 8.50 | Hurt |
| add attribute-conditioned editing before CW (StarGAN+CW) | CW · Face++ (T-FPR=1e-3) | attack success rate (%): 37.24 -> 47.45 | Improved |
| generate adversarial examples with SemanticAdv instead of PGD against adversarial training | adversarial-training defense · Adversarial-training defense | defense accuracy (%): 46.7 -> 10 | Improved |
| use SemanticAdv instead of CW on AMT perceptual realism | SemanticAdv · AMT user study, G-FPR<1e-4 | fraction judged reasonably-looking (%): 30.27 -> 39.14 | Improved |
Later papers that compared with this method
1 later accepted paper reports SemanticAdv as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- ColorFool: Semantic Adversarial Colorization (CVPR 2020 · better on 4 of 5 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Semantic Adversarial Attacks: Parametric Transformations That Fool Deep Classifiers (ICCV 2019 · similarity 0.92)
- Unrestricted Adversarial Examples via Semantic Manipulation (ICLR 2020 · similarity 0.90)
- Constructing Semantics-Aware Adversarial Examples with a Probabilistic Perspective (NeurIPS 2024 · similarity 0.89)
- PixelDefend: Leveraging Generative Models to Understand and Defend against Adversarial Examples (ICLR 2018 · similarity 0.89)
- Adversarial Examples for Semantic Segmentation and Object Detection (ICCV 2017 · similarity 0.89)
- Generative Adversarial Perturbations (CVPR 2018 · similarity 0.88)
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.
