Research map / Adversarial robustness and security / Attacks on text and graphs
NAG: Network for Adversary Generation
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 NAG instead of UAP | Universal adversarial perturbations for an ILSVRC-2014 target CNN · ILSVRC-2014 val, across the seven target CNNs | mean fooling rate: 57.66 -> 65.68 | Helped |
| Train one generator jointly against seven target CNNs | NAG · ILSVRC-2014 val, seven target CNNs | mean fooling rate: 72.62% -> 80.02% | Helped |
| Train the generator on an ensemble of target models rather than a single target | NAG ensemble generator · Left-out CaffeNet, VGG-19, and ResNet-152 | Mean BBFR: 71.26 -> 76.40 | Helped |
| Increase training data from 1000 to 50000 images (ablation) | NAG with ResNet-152 target · ILSVRC-2014 val | white-box fooling rate: 61.54 -> 91.16 | Helped |
| Increase training data from 1000 to 50000 images (ablation) | NAG with ResNet-152 target · ILSVRC-2014 val, black-box mean | Mean BBFR: 39.46 -> 67.45 | Helped |
| Increase diversity-loss weight from 0.1 to 10 (ablation) | NAG with ResNet-50 target · ILSVRC-2014 val | fooling rate: 89.04 -> 76.26 | Hurt |
| Increase diversity-loss weight from 0.1 to 10 (ablation) | NAG with ResNet-50 target · ILSVRC-2014 val | number of dominant labels: 253 -> 348 | Helped |
| Use Conv4 instead of softmax features for diversity (ablation) | NAG with VGG-F target · VGG-F held-out set | mean fooling rate: 93.57 (0.11) -> 93.86 (0.055) | No clear effect |
| Use Conv5 instead of softmax features for diversity (ablation) | NAG with VGG-F target · VGG-F held-out set | mean fooling rate: 93.57 (0.11) -> 94.80 (0.072) | No clear effect |
Closest papers by meaning
- Generative Adversarial Perturbations (CVPR 2018 · similarity 0.91)
- Adversarial vulnerability for any classifier (NeurIPS 2018 · similarity 0.91)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.89)
- Contrasting Adversarial Perturbations: The Space of Harmless Perturbations (AAAI 2025 · similarity 0.89)
- Detecting Adversarial Data Using Perturbation Forgery (CVPR 2025 · similarity 0.88)
- Theoretical Understanding of Learning from Adversarial Perturbations (ICLR 2024 · 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.