Research map / Image generation / Autoregressive and latent generation
ManiFPT: Defining and Analyzing Fingerprints of Generative Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Formalizes generative-model artifacts and fingerprints as deviations of generated samples from an estimated data manifold, and uses them as features for multi-class model attribution across GAN, VAE, flow and score-based models.
Problem
Fingerprints of generative models are used in practice (detection, attribution) but have no explicit definition, and prior work mostly studies binary real-vs-synthetic detection rather than multi-class attribution across many generative model families.
Approach
Define an artifact as the difference between a generated sample and its closest point on the estimated true data manifold (approximated by real training images in RGB, frequency, supervised or self-supervised embedding spaces), and define a model's fingerprint as the set of its artifacts; train a ResNet50 attributor on these artifacts.
Key results
- ManiFPT artifact features improve model attribution accuracy over the best baseline on all four GM datasets, by about 11.6%, 3.7% and 1.9% on three datasets and 5.73% on average (per the text).
- On cross-dataset generalization, ManiFPT variants reach up to 67.7% (CIFAR-10 to CelebA) versus 62.5% for the best baseline Wang20 in that scenario.
- Clustering analysis shows upsampling type and loss function types align best with the clustering structure of fingerprints (NMI 0.643 and 0.611 averages).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use artifact representations in supervised-learning feature space | ManiFPT model attribution · GM-CIFAR10 | accuracy (%): 62.01 -> 72.018 | Helped |
| use artifact representations in self-supervised feature space | ManiFPT model attribution · GM-CelebA | accuracy (%): 63.1 -> 74.7 | Helped |
| use artifact representations in frequency space | ManiFPT model attribution · GM-CHQ | accuracy (%): 61.1 -> 64.8 | Helped |
| use artifact representations in frequency space | ManiFPT model attribution · GM-FFHQ | accuracy (%): 64.2 -> 66.1 | Helped |
| vary the artifact embedding space among RGB, frequency, supervised-learning, and self-supervised-learning spaces (ablation) | ManiFPT model attribution · GM-CIFAR10, GM-CelebA, GM-CHQ, GM-FFHQ | accuracy (%): 69.48 ± 1.08; 70.191 ± 0.96; 72.018 ± 0.92; 70.177 ± 1.13; 70.5 ± 1.565; 72.8 ± 1.321; 73.6 ± 1.102; 74.7 ± 1.121; 63.7 | Mixed |
Surprising findings
- Hand-crafted color and frequency fingerprints perform worse than CNN-learned fingerprints for attribution.
- Attribution accuracy is lower on GM-CHQ than GM-CelebA, which the authors attribute to more advanced and hybrid models matching the data distribution better.
Limitations
- Attribution accuracy is lower for advanced and hybrid models, which are harder to attribute to a single source.
- Manifold is only estimated from real training images in a chosen embedding space, so the artifact depends on that choice.
Experimental setup
Datasets: CIFAR-10 CelebA-64 CelebA-HQ (256) FFHQ (256) GM-CIFAR10 GM-CelebA GM-CHQ GM-FFHQ
Benchmarks: GM-CIFAR10 GM-CelebA GM-CHQ GM-FFHQ
Metrics: accuracy Fréchet distance ratio (FDR) NMI
Baselines: McCloskey 2018 color cues Nataraj 2019 co-occurrence Durall 2020 Dzanic 2020 Wang 2020 Marra 2018 Marra 2019 Yu 2019
Builds on: Precision and Recall (Kynkäänniemi et al.) integral probability metrics Yu et al. 2019 GAN fingerprints Durall et al. 2020 frequency features Wang et al. 2020 CNN-generated image detection
Closest papers by meaning
- WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.87)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.84)
- How to Trace Latent Generative Model Generated Images without Artificial Watermark? (ICML 2024 · similarity 0.84)
- Single-Model Attribution of Generative Models Through Final-Layer Inversion (ICML 2024 · similarity 0.84)
- Position: All Current Generative Fidelity and Diversity Metrics are Flawed (ICML 2025 · similarity 0.84)
- Fingerprinting Denoising Diffusion Probabilistic Models (CVPR 2025 · similarity 0.84)
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.