Research map / Adversarial robustness and security / Attacks on text and graphs
Semantic Adversarial Attacks: Parametric Transformations That Fool Deep Classifiers
Ameya Joshi, Amitangshu Mukherjee, Soumik Sarkar, Chinmay Hegde
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 attributes A5 and A6 to single-attribute attack A1 (multi-attribute Fader Network) (ablation) | Adversarial Fader Network · CelebA gender classifier, 500 test images | target model accuracy (%): 52.0 -> 3.00 | Improved |
| add eyebrows attribute A9 to 5-attribute AttGAN attack (ablation) | Adversarial AttGAN · CelebA gender classifier, 500 test images | target model accuracy (%): 70.40 -> 39.40 | Improved |
| optimize attribute parameters instead of worst-of-10 random sampling (single attribute A1) | Adversarial Fader Network · CelebA gender classifier, 500 test images | target model accuracy (%): 89.00 -> 52.0 | Improved |
| optimize attribute parameters instead of random sampling (AttGAN, 5 attributes) | Adversarial AttGAN · CelebA gender classifier, 500 test images | target model accuracy (%): 32.80 -> 70.40 | Hurt |
| optimize attribute parameters instead of random sampling (AttGAN, 6 attributes) | Adversarial AttGAN · CelebA gender classifier, 500 test images | target model accuracy (%): 40.40 -> 39.40 | No clear effect |
| multi-attribute semantic attack versus PGD at equal l_inf bound | Multi-attribute semantic attack · CelebA gender classifier, pixel-space comparison (eps=1.74) | target model accuracy (%): 26.2 -> 1.00 | Improved |
| multi-attribute semantic attack versus FGSM | Multi-attribute semantic attack · CelebA gender classifier, pixel-space comparison (eps=1.74) | target model accuracy (%): 91.6 -> 1.00 | Improved |
| multi-attribute semantic attack versus Spatial attack | Multi-attribute semantic attack · CelebA gender classifier, pixel-space comparison (eps=1.74) | target model accuracy (%): 41.00 -> 1.00 | Improved |
| multi-attribute semantic attack versus CW-l_inf | Multi-attribute semantic attack · CelebA gender classifier, pixel-space comparison (eps=1.74) | target model accuracy (%): 1.00 -> 0.00 | Hurt |
| add attributes A5 and A6 to single-attribute attack A1 (ablation) | Adversarial Fader Network · CelebA age classifier (appendix) | target model accuracy (%): 70.0 -> 12.0 | Improved |
| increase dimensionality k of the semantic attack space (ablation) | Synthetic subspace and neural attacks · Synthetic MoG binary classifier | target model accuracy (%) | Improved |
| add ReLU nonlinearity to the transformation function (ablation) | Synthetic subspace and neural attacks · Synthetic MoG binary classifier | target model accuracy (%) | Hurt |
Closest papers by meaning
- SemanticAdv: Generating Adversarial Examples via Attribute-conditioned Image Editing (ECCV 2020 · similarity 0.92)
- Generative Adversarial Perturbations (CVPR 2018 · similarity 0.91)
- Constructing Semantics-Aware Adversarial Examples with a Probabilistic Perspective (NeurIPS 2024 · similarity 0.90)
- Generating Natural Adversarial Examples (ICLR 2018 · similarity 0.90)
- Exploring Semantic-constrained Adversarial Example with Instruction Uncertainty Reduction (NeurIPS 2025 · similarity 0.87)
- A Frustratingly Simple Yet Highly Effective Attack Baseline: Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5/4o/o1 (NeurIPS 2025 · similarity 0.87)
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.
