Research map / Adversarial robustness and security / Attacks on text and graphs
Generative Adversarial Perturbations
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 |
|---|---|---|---|
| replace iterative UAP construction with a trainable perturbation generator | ImageNet universal attacks · ImageNet validation, L2=2000 | fooling rate: 84.5% -> 93.9%; 79.5% -> 90.3%; 88.5% -> 94.9% | Helped |
| replace iterative UAP construction with a trainable perturbation generator | ImageNet universal attacks · ImageNet validation, L∞=10 | fooling rate: 78.8% -> 83.7%; 77.8% -> 80.1%; 78.9% -> 82.7% | Helped |
| increase the perturbation bound from L∞=7 to L∞=13 | GAP image-dependent attacks · ImageNet validation | fooling ratio: VGG-16: 66.9% -> 88.5%; VGG-19: 68.4% -> 90.7%; Inception-v3: 85.3% -> 99.5% | Helped |
| train against VGG-16 and VGG-19 jointly rather than only one model | GAP universal perturbations · ImageNet validation, transfer to ResNet-152 | fooling rate: 52.2% -> 54.1% (VGG-16 training); 49.0% -> 54.1% (VGG-19 training) | Helped |
| increase the perturbation bound from L∞=5 to L∞=20 | GAP targeted universal attacks · Cityscapes validation | success rate: 79.5% -> 97.2% | Helped |
| increase the perturbation bound from L∞=5 to L∞=20 | GAP non-targeted universal attacks · Cityscapes | mean IoU: 12.8% -> 2.1% | Hurt |
| replace a universal perturbation with image-dependent perturbations | GAP non-targeted segmentation attacks · Cityscapes, L∞=5 | mean IoU: 12.8% -> 6.9% | Hurt |
| use the ResNet Generator instead of U-Net (ablation) | GAP perturbation generators · Reported experiments | performance | Helped |
Closest papers by meaning
- NAG: Network for Adversary Generation (CVPR 2018 · similarity 0.91)
- Generate (Non-Software) Bugs to Fool Classifiers (AAAI 2020 · similarity 0.90)
- PixelDefend: Leveraging Generative Models to Understand and Defend against Adversarial Examples (ICLR 2018 · similarity 0.90)
- Defense Against Universal Adversarial Perturbations (CVPR 2018 · similarity 0.89)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.87)
- Detecting Adversarial Data Using Perturbation Forgery (CVPR 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.