acceptodds
Under review as a conference paper at ICLR 2027

Compositional Fingerprints for Open-Set AI-Generated Image Attribution

Abstract

AI-generated image attribution is increasingly important for digital forensics, copyright protection, and content accountability. In practice, however, attribution is inherently open-set: generative models evolve rapidly, and test images may come from generators unseen during training. This is particularly challenging because new generators often inherit architectures, reuse generation mechanisms, or share model lineages with known sources, causing their forensic traces to partially overlap. Existing methods typically represent each generator with a holistic fingerprint, making it difficult to distinguish shared traces from source specific organization. We propose compositional source fingerprints, based on the insight that generator identity can be characterized by how reusable latent generative patterns are combined rather than by a single indivisible fingerprint. We instantiate this idea with a Restricted Dynamic Hypergraph (RdH), in which fingerprint atoms represent reusable latent patterns and induce hyperedges among images. Restricted activation constructs compact, image-specific atom combinations, while degree-aware propagation suppresses broadly shared atoms with limited source selectivity. The resulting compositional representations support known-source attribution, unknown rejection, and novel source structure discovery. On five Cross-Era protocols, RdH improves open-set attribution AUROC over the strongest evaluated baselines by 9.89 percentage points, while also substantially improving clustering quality among unseen generators.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
I-generated image attribution, Open-set attribution, Compositional source fingerprints, Dynamic hypergraph

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xJ9dOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper tackles open-set AI-generated image attribution. The authors formulate the problem as determining which known generative model produces an image or rejecting it as being generated by an unseen generator. The central claim is that existing holistic fingerprint methods fail on unknown sources, as new generators reuse patterns from known ones rather than producing entirely novel fingerprints. The authors propose Restricted Dynamic Hypergraph, which models each image as activating a sparse combination of learnable fingerprint atoms.

Critical strengths

Open-set AI-generated image attribution is an important and timely problem that deserves further investigation. I appreciate the authors for choosing a meaningful research topic. However, I have serious concerns about the problem formulation and the overall modeling of the task.

The results reported in Table 1 are significantly higher than those of the existing state-of-the-art methods. This substantial performance gap raises concerns about the fairness and reliability of the comparisons.

Critical weaknesses

AI-generated image attribution is inherently an open-set problem. Redefining AI-generated image attribution as a novel-class discovery problem is not well motivated. The proposed formulation does not address a fundamentally new problem, but instead introduces an unnecessary reformulation of an existing open-set attribution task.

The claim that “existing attribution methods typically represent each source using a holistic fingerprint or a global source-discriminative representation” is largely speculative. The paper does not provide sufficient empirical evidence to establish that this is the primary reason why existing methods struggle with unseen generators.

The authors report results on only a single benchmark, which is insufficient to substantiate this claim.

The paper does not provide a comparison of computational complexity or runtime with existing methods.

The proposed method appears to be highly dependent on the extracted feature representations, yet the paper does not evaluate its robustness to different backbones.

Questions for authors

My main concern is that the authors' redefinition of open-set image attribution is not reasonable, and I am also not convinced by the proposed model. The authors seem to have relied too heavily on GPT in conducting the research. Overall, I find the entire paper provides very little meaningful insight to me.

Reviewer_TwE9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses open-set attribution of AI-generated images, where the goal is to attribute an image to a known generator or reject it as coming from an unknown source. Existing methods typically learn a holistic source fingerprint, which can obscure how forensic traces are shared or reorganized across generators. The paper proposes a compositional view: a new generator need not introduce an entirely new fingerprint but may reuse latent generative patterns ("atoms") from known sources and combine them differently. Based on this idea, the authors introduce RdH (Restricted Dynamic Hypergraph), which represents images as nodes and fingerprint atoms as hyperedges, learns compact image-specific atom combinations via restricted activation, and uses degree-aware propagation to reduce the influence of broadly shared atoms. Source prototypes are constructed from atom-combination features for known-source attribution and unknown rejection, and Novel Class Discovery (NCD) is used as an auxiliary analysis of unknown-source structure.

Critical strengths

The paper reframes open-set AI-generated image attribution from "learning a holistic source signature" to "modeling reusable forensic patterns and their combinations." It tries to address a real limitation of existing methods—closely related or evolving generators may share forensic traces, so a single holistic embedding can collapse distinct sources and obscure how shared evidence is organized. The experimental results are good.

Critical weaknesses
  1. The term "atom" is used pervasively but never precisely defined. The paper repeatedly refers to "fingerprint atoms" and "compositional atoms," yet it never specifies what an atom actually corresponds to in the feature space—whether it is a learned basis vector, a cluster centroid, a hyperedge prototype, or a direction in the CLIP embedding space. The paper characterizes atoms only through their behavioral role in the incidence matrix, and the spectral analysis in Figure 4 is explicitly described as a post-hoc statistical association rather than evidence of a one-to-one correspondence with generation mechanisms.

  2. The central assumption is not validated. The work is built on the premise that "a new generator need not introduce an entirely new fingerprint: it may reuse latent generative patterns already present in known sources, but combine them differently." This is an interpretive hypothesis about how generators share forensic traces, but the paper does not provide direct evidence that such reuse actually occurs at the level of identifiable generative patterns. The supporting evidence (atom overlap, masking curves) is behavioral and consistent with the hypothesis, but does not establish it. The authors should clarify whether this is a validated property or a working assumption, and discuss what evidence would falsify it.

  3. The discriminability of same-family generators with highly overlapping atoms is not examined. If two generators are derived from the same family and share almost the same atoms, can the proposed compositional representation still distinguish them? The paper reports high atom overlap between unknown and known sources (Table A.20, 7–10 out of 10), but does not report atom-combination similarity among unknown sources themselves, nor does it provide a same-family negative control. Without such an analysis, the claim that "source identity is reflected by the joint activation pattern" remains untested in the regime where it matters most—closely related generators.

  4. Batch dependence limits single-image applicability. RdH's propagated feature depends on the test batch (Section 3.3, Section 4.5, Appendix H.2), and the paper reports substantial degradation at small batch sizes. Real-world forensic scenarios often require immediate attribution of a single suspicious image, yet the paper does not report single-image (batch size 1) inference performance. The authors should either (i) report single-image results and characterize the degradation, (ii) provide a batch-independent variant (e.g., constructing incidence from a fixed reference set rather than the current batch), or (iii) explicitly state that the method is designed for batch-level attribution and adjust the claims accordingly.

Questions for authors
  1. The authors should either (i) provide a precise definition of what an atom represents and direct positive evidence supporting that characterization, or (ii) rename the concept to something more neutral (e.g., "latent compositional features") and adjust the claims accordingly.

  2. If two generators are derived from the same family and share almost the same atoms, can the proposed compositional representation still distinguish them?

  3. The authors should either (i) report single-image results and characterize the degradation, (ii) provide a batch-independent variant, or (iii) explicitly state that the method is designed for batch-level attribution and adjust the claims accordingly.

Reviewer_pG8UOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper represents a generator's fingerprint through combinations of reusable latent atoms. Its Restricted Dynamic Hypergraph learns sparse image-atom associations and limits propagation through widely shared atoms. The resulting features support known-source classification, unknown rejection, and exploratory clustering of unseen sources. Five Cross-Era settings form the main evaluation.

Critical strengths
  1. The compositional hypothesis leads to a concrete, testable design. Related generators may share individual traces without sharing their combinations. The atom-masking experiments in Section 4.3 and the component ablations in Table 3 support this account: removing restricted connections reduces AUROC from 92.84% to 85.52%, while removing degree-aware weighting substantially weakens unknown-source clustering. I find this evidence more persuasive than the architectural motivation alone. It gives the representation contribution substance.

  2. The main empirical improvement is large enough to matter. Table 1 reports 92.84% average AUROC, 9.89 percentage points above the strongest evaluated baseline, together with strong NMI/ARI results on unknown generators. The clustering result is useful corroboration because it asks whether the representation preserves distinctions among unseen sources, beyond separating known from unknown images. These results are the main reason I lean toward acceptance.

Critical weaknesses
  1. The interpretation of the Cross-Era gain remains partly confounded by dataset origin. Section 4.1 draws every known source from GenImage and every unknown source from AIGI-Holmes. Varying generator families across five protocols does not break this association. Collection or encoding differences could therefore contribute to unknown rejection, even when all methods receive the same training data. This does not establish that RdH exploits a shortcut, but it limits how confidently the 9.89-point gain can be attributed to compositional generator traces. The unknown-source clustering evidence helps, although it does not resolve this particular ambiguity.

  2. The joint result labeled OSCR has a narrower meaning than its prominence in Table 1 suggests. Section 4.1 states that it is accuracy at a median-midpoint threshold determined using the known/unknown test partition. The reported 84.50% is therefore a post-hoc operating-point result. The definition is disclosed, but it does not establish joint accuracy at a threshold chosen before seeing test membership. This concern affects that column; AUROC and the clustering results remain informative.

Questions for authors
  1. Clarification of the Cross-Era protocol: which content, acquisition, or image-encoding differences between GenImage and AIGI-Holmes remain after preprocessing? Which existing comparison best separates these effects from generator novelty? This would materially affect my confidence in the explanation of the AUROC gain.

  2. Evaluation clarification: is there an operating threshold selected without known/unknown test membership, and, if so, does the comparative joint-accuracy advantage remain at that threshold?

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

Thank you for your feedback. We summarize the additional work we carried out during the discussion.

In RdH, an atom is a learned vector in the same space as the image features. Different generators can activate the same atoms, and RdH uses their joint responses for attribution. To examine what these atoms capture, we changed one decoder operator at a time and compared reconstructions of the same image. We used a separate set of images to select the atoms and establish their expected response directions before evaluation. For example, when we replaced bilinear with nearest-neighbor upsampling, the joint response of five atoms shifted in the expected direction in 95% of paired comparisons on average, compared with 78% for the top-ranked atom alone. Across all four operator changes, individual atoms carried part of the evidence for the change, and their joint response captured it more consistently across decoder configurations (Appendix G.2, Table A.19).

We added four protocols to examine the effect of dataset origin on unknown-source rejection. In the original Cross-Era protocols, known sources came from GenImage and unknown sources from AIGI-Holmes. Two new protocols combined FLUX images from AIGI-Holmes and EvalGEN within the same unknown source. The other two swapped a known and an unknown generator so that GenImage and AIGI-Holmes each contributed both known and unknown sources. RdH achieved over 91% AUROC in each of the four protocols, although the swaps also changed the generator mix. We also examined how well RdH distinguished closely related generators. When SD3.5-Large was unknown, at most 0.20% of its test images were accepted as SD 1.5. When both were known, their bidirectional misclassification rate was 0.00% in the two swap protocols. The two unknown variants Janus-Pro-1B and Janus-Pro-7B proved harder to distinguish by their atom responses. They shared 27 of their 30 most active atoms, and cosine similarity was close to chance at determining whether two images came from the same variant (AUROC 0.55, Appendix G.1, Table A.17).

We also extended our analysis of test batches to single-query inference. RdH processes each new test image with 127 fixed reference images from known sources in the training set. With the same frozen checkpoints, mean AUROC across five protocols was 92.28% for individual queries and 92.63% for batches of 128 test images. On one H100 GPU, an individual query took 6.95 ms, compared with 1.09 ms per image averaged over a full batch. Batching was more efficient per image, while the fixed references allowed each query to be processed as soon as it arrived (Appendix H.2, Table A.21).

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.