Research map / Image generation / GANs
Attributing Fake Images to GANs: Learning and Analyzing GAN Fingerprints
Ning Yu, Larry S. Davis, Mario Fritz
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 |
|---|---|---|---|
| restrict to low-frequency band at downsample factor 8 (16x16) instead of full image (ablation) | attribution network · CelebA seed setup | accuracy (%): 99.14 -> 87.20 | Hurt |
| use high-frequency residual instead of low-frequency band at 16x16 (ablation) | attribution network · CelebA seed setup | accuracy (%): 87.20 -> 92.90 | Improved |
| use high-frequency residual instead of low-frequency band at 32x32 (ablation) | attribution network · LSUN seed setup | accuracy (%): 91.08 -> 96.04 | Improved |
| restrict to 4x4 low-frequency band (ablation) | attribution network · CelebA seed setup | accuracy (%): 99.14 -> 26.58 | Hurt |
| reduce patch size from 128x128 to 24x24 (ablation) | post-pooling network · CelebA seed setup | accuracy (%): 99.34 -> 99.24 | No clear effect |
| reduce patch size from 128x128 to 24x24 (ablation) | post-pooling network · LSUN seed setup | accuracy (%): 97.44 -> 88.36 | Hurt |
| reduce patch size to 3x3 (ablation) | post-pooling network · CelebA seed setup | accuracy (%): 99.34 -> 13.42 | Hurt |
| evaluate on the 10% artifact-free selected images instead of all images (ablation) | attribution network · CelebA seed setup | accuracy (%): 99.14 -> 99.93 | No clear effect |
| evaluate on the 10% artifact-free selected images instead of all images (ablation) | attribution network · LSUN seed setup | accuracy (%): 97.04 -> 98.16 | No clear effect |
| fine-tune classifier against noise attack (immunize) (ablation) | attribution network · CelebA noise attack | accuracy (%): 9.14 -> 93.02 | Improved |
| fine-tune classifier against JPEG compression attack (immunize) (ablation) | attribution network · CelebA compression attack | accuracy (%): 8.77 -> 88.02 | Improved |
| fine-tune classifier against combination attack (immunize) (ablation) | attribution network · CelebA combination attack | accuracy (%): 19.31 -> 72.64 | Improved |
Closest papers by meaning
- Towards Discovery and Attribution of Open-World GAN Generated Images (ICCV 2021 · similarity 0.93)
- RepMix: Representation Mixing for Robust Attribution of Synthesized Images (ECCV 2022 · similarity 0.92)
- Artificial Fingerprinting for Generative Models: Rooting Deepfake Attribution in Training Data (ICCV 2021 · similarity 0.91)
- Think Twice Before Detecting GAN-Generated Fake Images From Their Spectral Domain Imprints (CVPR 2022 · similarity 0.89)
- Deepfake Network Architecture Attribution (AAAI 2022 · similarity 0.89)
- Leveraging Frequency Analysis for Deep Fake Image Recognition (ICML 2020 · 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.
