acceptodds
Under review as a conference paper at ICLR 2027

Learning Generalizable Model Fingerprints for Attribution of Unseen Generative Models

Abstract

In this work, we propose a synthetic image attribution method based on generative model fingerprints that requires no synthetic images from real-world generative models during training. Instead, we train the fingerprint extractor entirely on a large and controllable bank of reconstruction models spanning diverse architectures and independently perturbed parameterizations, using the fingerprints embedded in their reconstructed images as supervisory signals. The proposed framework learns content-invariant fingerprints that distinguish not only different architectures but also different parameterizations of the same architecture, enabling direct attribution to previously unseen generative models. We further adopt an agent-guided collaborative training strategy to adaptively coordinate sampling, local evidence selection, loss configuration, frequency scheduling, and model-bank expansion during training. We evaluate the learned fingerprints on 26 unseen source models collected from four public benchmarks. Our method achieves a macro-average verification AUC of 97.01%, outperforming the strongest baseline by 5.46 percentage points, and shows clear separation between intra-source and inter-source fingerprints. These results demonstrate that fingerprints learned solely from controlled reconstruction models can generalize effectively to diverse unseen generative models.

Primary area:
interpretability and explainable AI
Keywords:
Synthetic Image Attribution, Model Fingerprinting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_EEsGOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses generative model attribution, which aims to identify the source model of a generated image. As the number of generative models continues to grow, collecting sufficient training data from every model becomes increasingly impractical. To address this challenge, the authors construct a proxy generative model using autoencoders with diverse architectures and initializations. Specifically, they use COCO images to generate outputs from these autoencoders and train a ResNet-based classifier to distinguish which autoencoder produced images. After training, they discard the classification head and reuse the encoder as a feature extractor. For attribution to actual generative models, they compute image embeddings and perform classification using model-specific prototype vectors. The key idea is to learn transferable model fingerprints from synthetic proxy models rather than directly training on a large collection of real generative models.

Critical strengths

The performance of the proposed paper, it achieved higher performance compared to FP

This paper proposed agent-driven training strategies in attribution field.

Critical weaknesses
  • The central idea highlighted in the abstract and the first contribution—training a fingerprint extractor solely on a bank of synthetic reconstruction models (288 architectures, 5,760 models) without real-generator images—appears to be the core contribution of Yang et al. (2025), which is used here only as the FP baseline and is not discussed in the related work or acknowledged in Sec. 3.2. The authors should explicitly attribute the model bank and the training paradigm to prior work, and reframe their contributions around what is actually new (the additional objectives, agent-guided scheduling, and bank expansion), supported by corresponding ablations.

  • Since the 26 sources are drawn from different benchmarks with likely differing resolutions, formats, and compression, and the method relies on residual FFT amplitudes, the reported separability may reflect dataset-level processing shortcuts rather than generator-specific fingerprints; cross-benchmark verification of the same model or evaluation under unified preprocessing is needed to rule this out.

Questions for authors
  • Explain the difference of FP, it seems that the core contribgution is entirely same as FP but this paper is neither explained in related work and not introduced in problem formulation. as FP uses autoencoders as scaleable proxy to determine its fingerprints, the authors should describe the difference

  • Explanations are poor, the authors should provide detailed setup

  • as stated earlier, the fingerprints can be learned through shortcuts, the reviewer strongly suggest the authors to experiments the analysis on fine-tuned model or post-processing. meaning, how does the fingerprints change if one fine-tunes its style of post-processes the image? Regarding the problem formulation, the fingerprints should remain fixed.

Reviewer_5hhHOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem of synthetic image attribution, i.e., identifying the generative models responsible for producing each synthetic image instance. In particular, the submission presents a pipeline for training a fingerprint extractor using images produced by reconstruction networks. By encouraging the learned fingerprints to be discriminative across different reconstruction network instances, the authors aim to learn fingerprint representations that can also distinguish the sources of synthetic images generated by previously unseen generative models.

Critical strengths
  • The submission addresses an important and well-established task of synthetic image attribution. It reports relatively strong performance, with decent improvements over existing baselines.
  • As far as I understand, the proposed training objective is reasonable, as it encourages the learned fingerprints to capture model-specific characteristics while suppressing variations caused by image content. By learning to distinguish a diverse set of reconstruction model instances, rather than directly relying on a fixed set of real-world generative models, the resulting representations may intuitively generalize better to previously unseen generative models.
Critical weaknesses
  • The presentation, particularly the method and experimental results sections, could be substantially improved by providing deeper insights and clearer motivations behind the design choices. In this regard, the current presentation remains relatively superficial, offering limited insights beyond the reported performance improvements. From my perspective, improving performance alone may not be sufficient to meet the standards of a top-tier ML conference, as I place greater value on scientific insights than on numerical improvements alone. The latter can be sensitive to experimental choices, and practices such as tuning hyperparameters on the test set or cherry-picking results can easily undermine the reliability of empirical findings.

  • Moreover, the key idea of using reconstruction networks instead of directly relying on real-world generative models appears to have already been explored in the published work of Fei et al. (2026). Although I am willing to consider the possibility that the two works were developed concurrently, this nonetheless somewhat lowers my assessment of the novelty of this submission, particularly given the limited new methodological insights provided.

Fei et al. (2026), "One for All: Synthesis-Free Fingerprint Learning for Attribution of In-the-Wild Synthetic Images", AAAI-26

Questions for authors
  1. How are the reconstruction architectures constructed? And how sensitive is the performance to these design choices, such as the diversity and number of reconstruction architectures and their parameterizations? Such details appear to be missing from the current manuscript.

  2. The proposed training procedure appears to rely on a careful weighting of different terms in the objective. However, there seems to be no experimental analysis of the sensitivity to these hyperparameters. How robust is the reported performance to different loss weight configurations?

  3. How does the proposed method compare with Fei et al. (2026), particularly given that both approaches rely on reconstruction-based fingerprint learning for attribution to unseen generative models? It would be helpful to clarify the key methodological differences and provide experimental evidence isolating the contributions of the newly introduced components.

Minor Points

  • There is inappropriate usage of \citet and \citep throughout the manuscript.
  • There are repeated author names, such as "Marra et al. Marra et al." and "Yu et al. Yu et al.", in the related work section. Similar citation formatting issues appear throughout this section.
Reviewer_7KcYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper tackles the problem of identifying which generative model has generated a given image. The main contribution is that this framework does not need training on outputs from the models that we want to detect. The proposed solution builds on top of One for All paper, with additional objective functions and a LLM guided training paradigm.

Critical strengths
  1. The paper has a strong experimental design, the evaluate across 20+ unseen models and show the effectiveness of the proposed method strongly. All 26 sources achieve above 90% AUC and show that there is no overfitting to specific architectures.
  2. The authors show ablation across all the proposed 4 objective functions. Although these are from existing known works, these loss functions address a problem that is explained intuitively and proven empirically.
Critical weaknesses
  1. Although one main contribution of the paper is LLM guided training, there is no ablation studies to show its impact. What happens if we remove that and just random sample or grid search? This weakens the effect of the proposed method. Similarly, what is the impact of the reconstruction bank, do the number of architectures impact generalization capability?
  2. Lack of important baselines: The paper builds on top of OneforAll, but does not provide any baseline comparisons with that work. Since that work also uses similar ideas, it is hard to understand the impact of proposed contributions.
  3. Practical impact: The introduction positions this paper as a IP protection and digital forensics, but the inference protoccol requires a pre-selected candidates of models. Additionally, the method requires 256 images per candidate model (no justification provided on number 256 and the impact of this choice).
Questions for authors
  1. Could you provide comparisons with OneforAll (Fei et al) to understand impact of proposed work?
  2. What is the difference between training with fixed hyperparameters and proposed method?
  3. What are the reconstruction architectures used and does the choice matter in results?
  4. Why do we need 256 images during inference per model? What happens if there are say 5 images?
  5. Are these 256 reference images from diverse prompts or from same categories/concepts? How much does this choice affect the performance?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.