Research map / Image generation / GANs
Artificial Fingerprinting for Generative Models: Rooting Deepfake Attribution in Training Data
Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, 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 |
|---|---|---|---|
| embed fingerprints in training data instead of jointly training the GAN to emit them | ProGAN · CelebA | FID: 60.28 -> 14.38 | Improved |
| use deep-learning steganography instead of conventional JPEG-based steganography | StyleGAN2 · CelebA | bitwise accuracy: 0.51 -> 0.99 | Improved |
| embed fingerprints in training data instead of jointly training the GAN to emit them | ProGAN · CelebA | bitwise accuracy: 0.93 -> 0.98 | Improved |
| fingerprint the training data | ProGAN · LSUN Cat | FID: 45.22 -> 48.97 | Hurt |
| fingerprint the training data | StyleGAN2 · CelebA | FID: 6.41 -> 6.23 | No clear effect |
| fingerprint the training data | BigGAN · CIFAR-10 | FID: 6.25 -> 6.80 | Hurt |
| add Gaussian noise (std 0.1) to generated images | fingerprinted ProGAN · CelebA | bitwise accuracy: 0.99 -> 0.77 | Hurt |
| apply model weight quantization (precision 10^0) | fingerprinted ProGAN · CelebA | bitwise accuracy: 0.99 -> 0.64 | Hurt |
| use artificial fingerprints instead of a CNN deepfake detector | ProGAN · CelebA deepfake detection | detection accuracy: 0.508 -> 1.000 | Improved |
| use artificial fingerprints instead of a CNN classifier for attribution | ProGAN · CelebA deepfake attribution | attribution accuracy: 0.235 -> 1.000 | Improved |
| give the attacker access to encoder and decoder with fingerprint codes 10 bits different | fingerprint secrecy · ATS attack | attacker classification accuracy: 0.503 -> 0.504 | No clear effect |
Closest papers by meaning
- Responsible Disclosure of Generative Models Using Scalable Fingerprinting (ICLR 2022 · similarity 0.91)
- ManiFPT: Defining and Analyzing Fingerprints of Generative Models (CVPR 2024 · similarity 0.90)
- TraceEvader: Making DeepFakes More Untraceable via Evading the Forgery Model Attribution (AAAI 2024 · similarity 0.88)
- D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.88)
- WildFake: A Large-Scale and Hierarchical Dataset for AI-Generated Images Detection (AAAI 2025 · similarity 0.87)
- WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models (CVPR 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.
