acceptodds
Withdrawn from ICLR 2027

DPG-Probe: Resolving the Semantic-Artifact Dilemma via Dual-Preserving Gradients for AIGI Forensics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The rapid evolution of Artificial Intelligence Generated Imagery presents severe forensic challenges characterized by two primary hurdles: the unprecedented fidelity and diversity of synthetic content, and the critical demand for robust, 'in-the-wild' real-world AIGI detection. To tackle these, current methods are largely based on advanced vision foundation models because they have already observed a vast visual world. However, this is precisely the fundamental reason why detection struggles to generalize. In the CLS space, real and fake images under the same paradigm theme are extremely close, meaning that when your test semantics deviate from the training set, the boundary structure moves with the object, rather than the intrinsic artifact. Addressing this limitation requires treating AIGI detection as a matter of provenance rather than standard semantic classification. Let represent a frozen VFM and a semantic-content and signal-structure preserving transformation (i.e., orthogonal rotations). While and should match on the core content of the image, any residual discrepancies are captured by hidden gradient directions () sensitive to . Therefore, we propose DPG-Probe, which leverages the residual discrepancies between and as a training-free, label-free probe. By taking a single backward pass of a self-consistency loss on intermediate CLS tokens through the frozen VFM, the resulting CLS gradients () capture a structural common pattern of synthetic image formation rather than surface content. Because the gradient probe relies on dual-preserving transformations, it suppresses high-level semantic shifts and maintains superior stability under wild degradation compared to pixel-level artifacts (like VAE-reconstruction). Comprehensive evaluations demonstrate that IPG-Probe significantly enhances cross-domain generalization and real-world robustness over prior detectors.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AIGI Detection, Multimedia Forensics

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

DPG-Probe does not directly use the image features extracted by the model to determine whether an image is real or fake. Instead, it observes how the model’s internal responses change after the same image is rotated. The authors believe that these changes can reflect differences in the generation sources of real and AI-generated images. Finally, the resulting features are compared with those of existing real and generated images using nearest-neighbor matching, in order to determine whether a new image is AI-generated.

Critical strengths
  1. The problem definition and narrative logic of the paper are relatively clear, and readers can quickly understand the generalization problem that the paper hopes to solve.
  2. Numerical experiments are sufficient and the argument angles are full, covering multiple dimensions such as semantic distribution, real open scenes, new generation models, image degradation, different gradient aggregation strategies, feature layer selection, kNN neighborhood parameters, and inference efficiency.
  3. The method does not require parameter training for the AIGI detection task, which reduces the need for model fine-tuning and re-training across generators, and the implementation process of the method is relatively straightforward.
Critical weaknesses

The paper reframes AIGI detection as a provenance discrimination problem rather than a semantic classification problem, but this perspective itself is not novel [Towards Universal Fake Image Detectors That Generalize Across Generative Models;RIGID: A Training-free and Model-Agnostic Framework for Robust AI-Generated Image Detection].

The paper assumes that gradient responses under dual-preserving transformations can reduce semantic information and highlight features related to the image source. However, this claim is mainly supported by PCA visualizations, gradient distributions, and limited cross-dataset similarity analysis. More theoretical analysis and stronger quantitative evidence are still needed.

The mechanism analysis in Figure 3 involves a certain degree of selection bias: the dimensions used to compare fake signatures across generators are first selected in a supervised manner based on the real/fake Fisher ratio. Therefore, the higher cross-dataset fake similarity cannot independently demonstrate that the gradient space inherently contains a generator-agnostic synthetic structure.

The assumption of “dual-preserving” is overly strong, and the authors do not actually verify this property. Although a 90° rotation usually does not change what objects are present in an image, this does not imply that the semantic representations inside the VFM remain unchanged. Therefore, a distinction should be made between image-level semantic identity preservation and representation-level semantic invariance.

Questions for authors

see weakness

Reviewer_Qo4LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DPG-Probe, a training-free framework for AI-generated image detection. It extracts gradient representations from a frozen vision foundation model using a multi-view rotation-consistency objective, followed by kNN classification with a labeled reference gallery. The authors report strong cross-domain detection performance across multiple benchmarks.

Critical strengths
  1. The paper evaluates the proposed method on a relatively broad range of benchmarks, including semantic-OOD and in-the-wild settings.
  2. The experimental section includes comparisons with existing methods and several ablation studies, providing empirical evidence of the proposed method's effectiveness.
Critical weaknesses
  1. Unclear connection between rotation-consistency gradients and image provenance: My main concern is the motivation for the proposed loss. It measures the CLS feature differences between an image and its rotated versions, but both real and AI-generated images undergo the same transformations. Even if rotations preserve semantic content and signal structure, why should the remaining feature differences systematically depend on whether an image was captured by a camera or generated by an AI model? What specific provenance-related signal is expected to produce different gradient directions for the two classes? The paper demonstrates that these gradients are discriminative, but does not sufficiently explain or verify why the observed differences reflect image provenance rather than the vision model's sensitivity to orientation, spatial layout, or other visual properties. Moreover, Figure 3 selects the top 48 dimensions using real/fake Fisher separation ratios, which require class labels. Therefore, the reported cross-dataset consistency is measured in a label-selected subspace, limiting its value as independent evidence that the gradients naturally capture shared provenance information.

  2. Unclear necessity of gradient extraction: For a single-layer MSE consistency loss, the resulting gradient representation is mathematically proportional to the direct difference between the original and rotated CLS features. Although the multi-layer formulation introduces additional Jacobian-weighted terms, the paper does not analyze their specific contribution or compare against directly using forward feature differences. It is therefore unclear what additional information the backward pass provides.

  3. Computational overhead and reference dependence: Despite being training-free, the method requires backward propagation and a large labeled reference gallery for kNN classification, resulting in substantial inference overhead. These limitations reduce its practical efficiency and should be more clearly distinguished from the advantages of training-free representation extraction.

Questions for authors

Please refer to the Critical Weaknesses.

Reviewer_pL5TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DPG-Probe computes multi-layer CLS consistency losses across four orthogonal rotations through a frozen MetaCLIP-2 encoder. A backward pass produces CLS-state gradients, which are differentially aggregated and normalized for kNN detection using approximately 324,000 labeled SD v1.4 reference images. The paper evaluates cross-domain detection, corruption robustness, and inference costs. It performs no parameter optimization, but detection requires reference labels and substantial computation.

Critical strengths
  1. There are substantial reported benefits on external two-class datasets beyond improvements over CLS-kNN. On Chameleon, real/fake accuracy is .962/.937, averaging .950 versus DDA’s .824; on CommunityAI, it is .965/.978, averaging .971 versus DDA’s .847 (Table 1). These results retain balanced class performance and do not depend on the mixed aggregate metrics.

  2. The same-backbone/reference-bank comparisons in Tables 3–4 support investigating the rotation-gradient family. The appendix also includes controlled corruptions, layer/k sensitivity, and explicit cost measurements.

Critical weaknesses
  1. The ablations do not sufficiently isolate the contribution or justify its computation. FUNGI [A] already uses self-supervised gradients with kNN, and ConV [B] uses transformation consistency for generated-image detection. DPG’s state-gradient construction differs, but needs a same-view forward rotation-residual baseline and single-layer versus multi-layer loss controls. The single-layer MSE case reduces to a forward residual direction, although this does not establish equivalence for the actual multi-layer method. PatchQuad changes spatial content as well as the transformation, so it cannot isolate semantic preservation. Anchor is already close to the full method (.840/.938 versus .845/.942 in Tables 3–4), and the aggregation variants lack explicit definitions. Appendix A reports approximately 358 ms/image and 17.5 GB on an L40, yet provides no two-view versus four-view accuracy/latency/memory ablation. Batch-size profiling does not answer this question. Deferring view pruning to future work does not establish the necessity of the current design, and the existing comparisons do not establish semantic purification.
  2. Evaluation definitions and reporting need reconciliation. Tables 1–2 mix fake-only recall with two-class accuracy. Table 2 reports UnivFD .756 on AIGI-Now pixel, whereas the corresponding Table 19 entries all lie between .475 and .631; Appendix E’s dataset roster also differs from the main tables. WebAIG-25’s August–September 2025 dates do not substantiate its stated post-2025 exclusion from every VFM’s pretraining. These issues limit confidence in aggregate claims without directly invalidating the Chameleon/CommunityAI results. Generalization is also uneven: WildRF real accuracy is .475, and the mean .729 is below DDA’s .904 (Table 1). Comparisons at common real-image false-positive rates would clarify the practical scope.

Reference

[A] Simoncini, W., Gidaris, S., Bursuc, A., & Asano, Y. M. (2024). No train, all gain: Self-supervised gradients improve deep frozen representations. Advances in Neural Information Processing Systems, 37.

[B] Zhang, Y., Nie, J., Tian, X., Gong, M., Zhang, K., & Han, B. (2025). Detecting generated images by fitting natural image distributions. Advances in Neural Information Processing Systems, 38.

Questions for authors
  1. Under the same backbone, reference bank, and source-only selection protocol, how do two-view and four-view variants compare in two-class detection performance, latency, and peak memory? This is needed to justify the current design independently of future acceleration work.
  2. Please define Anchor and AbsMean mathematically and compare normalized forward rotation residuals with the proposed gradients, including single-layer versus multi-layer losses. What evidence supports differential purification being essential given the small gain over Anchor?
  3. Can you reconcile Tables 2 and 19, provide a consistent dataset/split/metric manifest, and explain how layer/k were selected? Please distinguish fake-only recall from two-class metrics and document or narrow WebAIG-25’s leakage claim.

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

Related papers

Discussion (0)

Sign in to comment.