Research map / Adversarial robustness and security / Certified and empirical robustness
Constructing Unrestricted Adversarial Examples with Generative Models
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 |
|---|---|---|---|
| Synthesize class-conditioned examples from an AC-GAN and optimize latent codes to target classifier predictions | Adversarially trained classifiers · MNIST, SVHN, CelebA | targeted attack success rate (%): 85.2; 84.2; 91.1 | Helped |
| Add trainable bounded image noise (ablation) | Unrestricted targeted attack · SVHN, adversarially trained ResNet | targeted attack success rate (%): 84.2 -> 91.6 | Helped |
| Add trainable bounded image noise (ablation) | Unrestricted targeted attack · CelebA, adversarially trained ResNet | targeted attack success rate (%): 91.1 -> 86.7 | Hurt |
| Add trainable bounded image noise (ablation) | Unrestricted targeted attack · MNIST, Madry network | targeted attack success rate (%): 85.2 -> 85.0 | No clear effect |
| Use unrestricted synthesis rather than small-norm PGD perturbations | Certified classifiers · MNIST | untargeted attack success rate (%): 86.6; 88.8 | Helped |
| Generate adversarial examples from scratch rather than use large PGD perturbations | Certified classifiers · MNIST, A/B human identification test | correct identification rate (%): 92.9% vs 76.8%; 87.6% vs 78.2% | Helped |
| Add trainable bounded image noise (ablation) | Unrestricted attack targeting Madry Net · MNIST transfer to other classifiers | classifier accuracy (%): 95.1 -> 78.3; 92.7 -> 73.8; 93.7 -> 84.9; 77.1 -> 78.1; 84.3 -> 63.0 | Mixed |
| Condition the generator on the intended source class and verify generated labels with human judgments | Unrestricted adversarial examples · MNIST, SVHN, CelebA | legitimacy and source-class validity | Helped |
Closest papers by meaning
- GAT: Generative Adversarial Training for Adversarial Example Detection and Robust Classification (ICLR 2020 · similarity 0.91)
- A Direct Approach to Robust Deep Learning Using Adversarial Networks (ICLR 2019 · similarity 0.91)
- One Man's Trash Is Another Man's Treasure: Resisting Adversarial Examples by Adversarial Examples (CVPR 2020 · similarity 0.91)
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.90)
- Generating Natural Adversarial Examples (ICLR 2018 · similarity 0.90)
- Adversarial Machine Learning at Scale (ICLR 2017 · 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.