Introduction
The rapid advancement of generative models (GMs) like Stable Diffusion and Midjourney has created a pressing need for reliable model attribution—identifying the specific AI architecture responsible for creating a piece of content. Beyond digital forensics, this is crucial for mitigating "model collapse," where AI performance degrades after being trained on synthetic data. Current state-of-the-art methods, such as ManiFPT, hypothesize that these models leave unique "fingerprints" as artifacts or deviations from the manifold of real data. However, these approaches typically assume a flat Euclidean geometry. Because high-dimensional data like natural images actually reside on complex, curved surfaces, Euclidean metrics often fail to accurately capture these subtle geometric signatures, especially when models or datasets change.
Methodology: From Euclidean to Riemannian Geometry
We propose a framework that leverages Riemannian geometry to define and extract model-specific fingerprints. The approach consists of three primary stages:
1. Learning the Manifold and the Pullback Metric
Instead of assuming a flat space, we use a Variational Autoencoder (VAE) to learn the latent geometry of real data. To measure distances accurately on this curved surface, we employ a "pullback metric." This metric essentially "pulls" the distances from the high-dimensional image space back into the latent space, providing a mathematically rigorous way to calculate lengths and distances on the manifold.
2. Defining the Artifact via Riemannian Center of Mass (RCM)
The core of our method is the definition of a "geometric artifact"—the difference between a generated sample and its projection onto the real data manifold. Traditional methods use Euclidean nearest-neighbor searches, which often result in projections that "float" in empty space, far from the actual data distribution. We replace this with the Riemannian Center of Mass (RCM), also known as the Fréchet mean.
As illustrated in the figure above, the RCM ensures that the projection remains strictly on the manifold. This results in an artifact vector that more cleanly isolates the intrinsic computational traces of the generative model rather than noise from off-manifold projections.
3. Model Attribution
These refined geometric artifacts are then processed by a standard ResNet-50 classifier. The classifier is trained to recognize the specific "geometric signature" of 27 different model architectures, ranging from GANs and VAEs to modern Diffusion models.
Experimental Results
The proposed R-GMfpts (RCM) method was evaluated against extensive benchmarks, including a new multimodal extension featuring vision-language models like Flux.1 and DALL-E 3.
- Superior Accuracy: On the GM-CIFAR10 dataset, our method achieved 78.17% accuracy, significantly outperforming the Euclidean-based ManiFPT (70.19%) and traditional frequency-based baselines.
- Enhanced Generalization: A key strength of the Riemannian approach is its ability to generalize to unseen data. When trained on CIFAR-10 but tested on CelebA, the method maintained high performance (up to 67.8% accuracy), whereas Euclidean methods struggled to identify signatures across different subject matters.
- Distinct Fingerprints: Using the Fréchet Distance Ratio (FDR), we observed that Riemannian artifacts form much tighter and more separable clusters for different model architectures, providing a more reliable "signature" for forensics.
Conclusion
By moving from flat Euclidean rulers to curved Riemannian metrics, we can more effectively isolate the unique geometric flaws inherent in generative model architectures. This approach provides a robust, theoretically grounded tool for AI forensics that generalizes across datasets and modalities. Future research will address the computational overhead of RCM calculations and investigate the framework's resilience to adversarial attempts to mask these geometric fingerprints.
Summary updated