Research map / Adversarial robustness and security / Certified and empirical robustness
On the Robustness of Quality Measures for GANs
Motasem Alfarra, Juan C. Pérez, Anna Frühstück, Philip H. S. Torr, Peter Wonka, Bernard Ghanem
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 |
|---|---|---|---|
| replace nominal Inception with robust Inception (kappa=128) in FID (ablation) | FID · ImageNet pixel attack (eps=0.01) | FID under attack: 7929.01 -> 4.5 | Improved |
| use larger adversarial training radius kappa=128 instead of 64 for the robust Inception (ablation) | R-FID · ImageNet pixel attack (eps=0.03) | R-FID under attack: 657.9 -> 264.6 | Improved |
| use larger adversarial training radius kappa=128 instead of 64 (ablation) | R-FID · CIFAR10 pixel attack (eps=0.03) | R-FID under attack: 46.4 -> 19.7 | Improved |
| replace nominal Inception with robust Inception (kappa=128) in IS (R-IS) (ablation) | Inception Score · CIFAR10 pixel attack (eps=5e-3) | IS under attack: IS drops from 11.54 to 2.62 vs R-IS drops from 9.94 to 5.49 | Improved |
| replace nominal Inception with robust Inception in IS against noise-image attack (ablation) | Inception Score · CIFAR10 random noise images | IS: 94.87 -> 1.01 | Improved |
| replace nominal Inception with robust Inception (kappa=128) against noise-image attack (ablation) | FID · ImageNet optimized noise images | FID: 1.05 -> 340 | Improved |
| replace nominal Inception with robust Inception (kappa=128) under latent w-space attack (ablation) | FID on StyleGANv2 · StyleGANv2 FFHQ truncation 1.0 (w-space attack) | metric increase under attack: 113.8 -> 114.3 compared to 2.54 -> 6.4 | Improved |
| replace nominal Inception with robust Inception (kappa=64) under z-space latent attack (ablation) | FID on StyleGANv2 · StyleGANv2 FFHQ truncation 0.7 (z-space attack) | relative metric increase: R-FID increases by 44% vs FID increase of 1000% | Improved |
| use robust Inception to measure StyleGANv2 vs FFHQ distance (ablation) | FID · StyleGANv2 FFHQ truncation 1.0 | FID vs R-FID: 2.6 -> 113.8 | Improved |
| lower StyleGANv2 truncation alpha from 1.0 to 0.7 (better visual quality) (ablation) | standard FID · StyleGANv2 FFHQ | FID: 2.65 -> 21.81 | Hurt |
| latent z-space attack on StyleGANv2 maximizing FID (alpha=1.0) | StyleGANv2 with FID · StyleGANv2 FFHQ | FID: 2.65 to 31.68 | Hurt |
| increase Gaussian noise sigma on ImageNet from 0.1 to 0.4 | R-FID kappa=128 · ImageNet degraded | R-FID: 16.65 -> 198.3 | Improved |
Closest papers by meaning
- Quality Aware Generative Adversarial Networks (NeurIPS 2019 · similarity 0.90)
- GREAT Score: Global Robustness Evaluation of Adversarial Perturbation using Generative Models (NeurIPS 2024 · similarity 0.89)
- Improved Techniques for Training Score-Based Generative Models (NeurIPS 2020 · similarity 0.89)
- Comparing the Robustness of Modern No-Reference Image- and Video-Quality Metrics to Adversarial Attacks (AAAI 2024 · similarity 0.88)
- Adversarial Score Distillation: When score distillation meets GAN (CVPR 2024 · similarity 0.87)
- ImageNet-D: Benchmarking Neural Network Robustness on Diffusion Synthetic Object (CVPR 2024 · similarity 0.86)
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.
