Research map / Image generation / Autoregressive and latent generation
Responsible Disclosure of Generative Models Using Scalable Fingerprinting
Ning Yu, Vladislav Skripniuk, Dingfan Chen, 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 |
|---|---|---|---|
| Use filter-modulated fingerprints rather than passing fingerprints through the generator input (ablation) | StyleGAN2 fingerprinting · LSUN Cat | bit accuracy: 0.500 -> 0.996 | Improved |
| Modulate filters separately with latent and fingerprint embeddings (ablation) | StyleGAN2 fingerprinting · LSUN Cat | bit accuracy: 0.996 -> 0.869 | Hurt |
| Modulate filters using concatenated latent and fingerprint embeddings (ablation) | StyleGAN2 fingerprinting · LSUN Bedroom | FID: 20.50 -> 64.91 | Hurt |
| Modulate all generator layers instead of a single 16×16 layer (ablation) | StyleGAN2 fingerprinting · CelebA | bit accuracy: 0.993 -> 0.991 | No clear effect |
| Modulate all generator layers instead of a single 8×8 layer (ablation) | StyleGAN2 fingerprinting · LSUN Bedroom | bit accuracy: 0.950 -> 0.993 | Improved |
| Add the consistency loss (ablation) | Fingerprint GAN with adversarial and fingerprint reconstruction losses · CelebA | FID: 12.84 -> 12.24 | Improved |
| Add latent-code reconstruction loss (ablation) | Fingerprint GAN with adversarial, fingerprint, and consistency losses · CelebA | FID: 12.24 -> 11.50 | Improved |
| Use only adversarial and fingerprint reconstruction losses, omitting consistency and latent reconstruction (ablation) | Fingerprint GAN · CelebA | bit accuracy: 0.999 vs 0.991 | No clear effect |
| Use on-the-fly fingerprint sampling instead of a finite training fingerprint set (ablation) | StyleGAN2 fingerprinting · CelebA | testing bit accuracy: 0.752 -> 0.991 | Improved |
| Train with a finite set of 10 fingerprints rather than sampling fingerprints on the fly (ablation) | StyleGAN2 fingerprinting · CelebA | testing bit accuracy: 0.991 -> 0.512 | Hurt |
| Train with perturbed-image augmentation for immunization (ablation) | Fingerprint GAN · CelebA under cropping, Gaussian noise, and combined perturbations | bit accuracy | Improved |
| Use proactive fingerprint attribution instead of passive classifiers | Deep fake detection and attribution · CelebA open world, 10 GAN sources | accuracy: 0.102 -> 1.000 | Improved |
Closest papers by meaning
- ManiFPT: Defining and Analyzing Fingerprints of Generative Models (CVPR 2024 · similarity 0.89)
- Scalable Fingerprinting of Large Language Models (NeurIPS 2025 · similarity 0.89)
- WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.89)
- Queries, Representation & Detection: The Next 100 Model Fingerprinting Schemes (AAAI 2025 · similarity 0.87)
- Fingerprinting Denoising Diffusion Probabilistic Models (CVPR 2025 · similarity 0.86)
- OmniMark: Efficient and Scalable Latent Diffusion Model Fingerprinting (AAAI 2025 · 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.
