acceptodds
Withdrawn from ICLR 2027

Efficient Black-Box Fingerprinting For Image Generation Models With Ambiguous Images

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Black-box attribution of image generation models is difficult because fine-tuning and other modifications can substantially alter their outputs. Text-only fingerprints are fragile because semantic preferences become entangled with unconstrained variation in style, layout, and composition. We present AmbiPrint, a passive framework that probes how models resolve controlled semantic ambiguity. Am- biPrint pairs a neutral text prompt with an ambiguous image that guides genera- tion and constrains visual variation without determining the final class. Across 25 open- and closed-source models, AmbiPrint separates source derivatives from in- dependently trained models under fine-tuning, compression, and substantial style and domain shifts. It succeeds with only 100 generations per model where text- only baselines fail. Adaptive attacks can evade attribution by targeting every eval- uated category, but doing so substantially reduces output diversity. AmbiPrint provides a practical and robust signal for black-box model attribution.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
diffusion, fingerprinting, watermarking, backdoor, dense

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_NYaCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes AmbiPrint for attributing a queried image-generation model to a protected source family. The method combines ambiguous image probes with neutral text, generates multiple outputs, and averages a pretrained classifier's responses to form a model fingerprint. Attribution is based on the distance between fingerprints. Experiments on 25 models report separation between derivatives and independent models, including under fine-tuning, compression, and domain shifts, with fewer generations than the evaluated text-only baselines.

Critical strengths
  1. Identifying model lineage using only query access is a meaningful problem. The evaluation includes several kinds of model modification and attempts to distinguish derivatives from independently trained models.
  2. The reported improvement over text-only fingerprints is substantial.
Critical weaknesses

The experiments do not establish why the proposed semantic probing is necessary compared with conventional image-forensic fingerprints. Averaging classifier outputs does not rule out sensitivity to model-specific low-level artifacts. Prior work has shown that generated images contain useful forensic signals [1], and fingerprint-based methods have also investigated relationships between fine-tuned generators [2]. The comparisons with CSF and text-only AmbiPrint establish the benefit of image conditioning relative to those baselines, but do not show an advantage over artifact-based attribution using a collection of generated images. This leaves both the semantic explanation and the practical value of the proposed design insufficiently supported.

[1] Wang et al., "CNN-generated images are surprisingly easy to spot... for now", CVPR 2020. [2] Sinitsa et al., "Deep Image Fingerprint: Towards Low Budget Synthetic Image Detection and Model Lineage Analysis".

Questions for authors
  1. Could the authors compare AmbiPrint with representative image-forensic fingerprint methods under the same model-family splits and generation budgets? A simple baseline that aggregates frequency-domain statistics or residual fingerprints across generated images would also be informative. Please evaluate both derivative detection and false attribution of independent models. The question is whether AmbiPrint provides an advantage in accuracy, robustness, or sample efficiency under comparable access assumptions.
  2. What evidence distinguishes a semantic preference signal from low-level artifacts preserved in the classifier outputs? One possible test is to apply semantics-preserving rotations or affine transformations to the generated outputs before attribution and compare the degradation of AmbiPrint and forensic baselines. If AmbiPrint remains reliable where the baselines degrade, this would establish a concrete advantage. Transformation robustness alone would not prove the absence of low-level information, since some forensic traces can survive these operations. Other controlled experiments that demonstrate the claimed advantage would also address this concern; this particular ablation is a suggestion, not a required implementation.
Reviewer_RZibOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The objective of this paper is to create means to tell if a set of image generation models are rooted from the same source model or not.

Their method uses a set of (image, text) pairs where the image is some average of images from the same category and the text is a description of that category. The underlying research hypothesis (I guess) is that different source models have different generation dynamics on such probing pairs, allowing models to be classified in the output space.

While this idea is interesting and worth exploring, my major concern is the lack of mechanistic understanding of the research hypothesis. Specifically, why this type of probes? Would other images, not of an averaging nature, and other text prompts work? how many probes do you need in theory (for each generative models)? Ideally we need provable sufficient conditions on attributability.

There are other technical concerns: What if we fine-tune a copied model against this fingerprint? The authors also explicitly mentioned issues with held-out models with similar generation dynamics, without addressing the issue.

Overall, an interesting yet raw idea that deserves more investigation.

Critical strengths

The problem statement is clean. The solution is easy to follow.

Critical weaknesses

The claims are purely empirical. The authors mentioned existing issues with held-out models.

Questions for authors

n/a

Reviewer_LPbNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes AmbiPrint, a passive black-box framework for source attribution of image-generation models. It combines supercategory text prompts with images constructed by mixing multiple classes, preserving semantic ambiguity while limiting visual variation unrelated to the model's origin. Generated images are projected through a pretrained classifier or encoder, and their averaged responses form fingerprints. Attribution is performed by comparing fingerprint distances against a calibrated threshold.

Critical strengths
  1. The paper is well-written, and the proposed fingerprinting pipeline is easy to follow.
  2. The method is evaluated across several model modifications, providing useful evidence of its applicability.
Critical weaknesses
  1. My main concern is whether the zero-error result with only 100 generated images per supercategory is stable. Would you obtain similar results with a different set of test images and repeated generation runs? Also, was λ = 0.2 in Section 4.3 chosen before the final tests, or after examining those test results? Please explain how you selected this parameter and provide results from independent repetitions.

  2. Table 1 shows that AmbiPrint performs better than CSF, but it is unclear how CSF was implemented. Section 4.4 allows it to select a favorable threshold, but did you also use its complete original processing and decision procedure? Please explain the implementation so we can assess the fairness of the comparison.

Questions for authors
  1. With 100 generations per supercategory, what are the false-positive and false-negative rates across independent runs using newly sampled probe images and fresh generations?
  2. Did the CSF implementation reproduce the original prompt construction, output processing, and attribution procedure?

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

Related papers

Discussion (0)

Sign in to comment.