Riemannian-Geometric Fingerprints of Generative Models

Hae Jin Song, Laurent Itti

TL;DR

This paper introduces a novel framework for "fingerprinting" generative models (GMs) by leveraging Riemannian geometry. By treating data as lying on a curved manifold and using learned geometric distances, they achieve significantly improved accuracy in identifying the source model of generated content, even across different datasets and model types.

What Problem Does This Paper Address?

The rapid proliferation of generative models (like Stable Diffusion for images and GPT for text) has created a pressing need for methods to identify the source of generated content. This is crucial for intellectual property protection, combating misinformation, and understanding model collapse (where models trained on synthetic data degrade). While previous work has observed "fingerprints" or artifacts left by GMs, these methods often assume data lies on a flat, Euclidean space, which is unrealistic for complex data like images and text. This paper addresses the lack of a principled, generalizable framework for defining and detecting these fingerprints, especially for non-Euclidean data.

Methods

The core innovation is the application of Riemannian geometry to model attribution. Instead of treating data as points in a flat space, the authors model the data distribution as residing on a curved manifold.

Here's a breakdown of their approach:

  • Learning the Data Manifold: They use a Variational Autoencoder (VAE) trained on real data to learn a latent space representation. The VAE's decoder is then used to "pull back" the geometric structure (metric) from the observed data space (e.g., pixel space) to this latent manifold. This effectively creates a learned Riemannian metric for the latent space.
  • Defining Artifacts: An "artifact" of a generative model is defined as the deviation of a generated sample from the learned real data manifold. This is quantified by projecting the generated sample onto the manifold.
  • Riemannian Center of Mass (RCM): To find this projection, they employ a Riemannian Center of Mass (RCM) approach. Instead of simple Euclidean distance, they use geodesic distances (the shortest paths on the curved manifold) to find the k-nearest neighbors of a generated sample in the latent space and then compute their RCM. This RCM serves as the projection onto the manifold.
  • Fingerprint: The fingerprint of a generative model is the collection of all such artifacts from its generated samples.
  • Attribution Network: A ResNet classifier is trained on these artifact features to predict the source generative model.

Datasets

The authors evaluate their method on an extended benchmark dataset that includes:

  • Real Datasets: CIFAR-10, CelebA, CelebA-HQ, and FFHQ.
  • Generative Models: A large array of models from GAN, VAE, Flow, and Score-based families, totaling 27 architectures.
  • Resolutions: 64x64 and 256x256.
  • Modalities: Vision and Vision-Language models (text-to-image).

They emphasize that their dataset is designed to ensure consistency by training attribution methods on models that used the same real dataset, allowing for independent evaluation of model architecture and dataset effects.

Key Insights and Ideas

  • Data is Not Flat: The fundamental insight is that real-world data distributions are rarely Euclidean. Ignoring this curvature leads to less effective fingerprinting.
  • Geometry as a Fingerprint: The geometric properties of the data manifold, learned from real data, provide a powerful signal for distinguishing between different generative models.
  • RCM for Accurate Projection: Using a Riemannian Center of Mass, informed by learned geodesic distances, provides a more accurate projection of generated samples onto the true data manifold compared to simpler Euclidean methods. This leads to a more precise artifact measurement.
  • Generalization is Key: The proposed method shows strong generalization capabilities, not just across different datasets but also across different model types and even modalities, suggesting its practical utility.

Key Results

  • Superior Attribution Accuracy: The proposed Riemannian-geometric fingerprinting method (R-GMfptSRCM) consistently outperforms all baseline methods, including Euclidean-based manifold methods, across all tested datasets. For example, on GM-CIFAR10, their method achieved 78.17% accuracy compared to the best baseline of 70.19%.
  • Improved Separability: The method also shows significantly higher Fréchet Distance Ratios (FDR), indicating that the learned fingerprints are more distinct and easier to separate between different models.
  • Strong Generalization: The method demonstrates robust generalization across datasets with different semantic domains (e.g., CIFAR-10 vs. CelebA) and across different model architectures and modalities.
  • Robustness: The method is shown to be robust against common post-processing techniques like Gaussian blurring and JPEG compression.

Limitations and Future Work

  • Computational Cost: While the paper presents a gradient-based algorithm for computing RCM, the process of learning the Riemannian manifold and computing fingerprints can be computationally intensive.
  • Assumption of Manifold Hypothesis: The method relies on the assumption that data lies on a lower-dimensional manifold, which is generally true but might have edge cases.
  • Future Work: The authors suggest extending their framework to other modalities like audio and video, and exploring its application in detecting model collapse and understanding the internal workings of GMs. They also highlight the potential for further research into cross-modal fingerprinting.

Summary updated

Abstract

Recent breakthroughs and rapid integration of generative models (GMs) have sparked interest in the problem of model attribution and their fingerprints. For instance, service providers need reliable methods of authenticating their models to protect their IP, while users and law enforcement seek to verify the source of generated content for accountability and trust. In addition, a growing threat of model collapse is arising, as more model-generated data are being fed back into sources (e.g., YouTube) that are often harvested for training (“regurgitative training”), heightening the need to differentiate synthetic from human data. Yet, a gap still exists in understanding generative models’ fingerprints, we believe, stemming from the lack of a formal framework that can define, represent, and analyze the fingerprints in a principled way. To address this gap, we take a geometric approach and propose a new definition of artifact and fingerprint of GMs using Riemannian geometry, which allows us to leverage the rich theory of differential geometry. Our new definition generalizes previous work (Song et al., 2024) to non-Euclidean manifolds by learning Riemannian metrics from data and replacing the Euclidean distances and nearest-neighbor search with geodesic distances and kNN-based Riemannian center of mass. We apply our theory to a new gradient-based algorithm for computing the fingerprints in practice. Results show that it is more effective in distinguishing a large array of GMs, spanning across 4 different datasets in 2 different resolutions (64×64, 256×256), 27 model architectures, and 2 modalities (Vision, Vision-Language). Using our proposed definition significantly improves the performance on model attribution, as well as a generalization to unseen datasets, model types, and modalities, suggesting its practical efficacy.
BibTeX
@inproceedings{song2025riemannian,
  author = {Hae Jin Song and Laurent Itti},
  title = {{Riemannian-Geometric Fingerprints of Generative Models}},
  booktitle = {BEW @ ICCV},
  year = {2025},
  url = {https://openreview.net/forum?id=00LViVLqkN},
}
Figure