acceptodds
Under review as a conference paper at ICLR 2027

EditInk: Learning Edit-Sensitive Ink to Make AI Manipulation Legible

Abstract

Generative image editing can produce visually convincing local modifications, making manipulated regions difficult to identify. Proactive localization addresses this challenge by embedding signals before image release to provide additional localization cues. However, jointly learning a signal and its reader from actual editing outcomes remains challenging when closed-source editors allow forward queries but not backpropagation. To address this challenge, we propose a localization framework named EditInk that learns a shared edit-sensitive ink together with an ink developer. The ink is an image-embedded signal learned to remain imperceptible before editing and provide spatial cues when disrupted, while the developer reads these cues to identify manipulated regions from the edited image alone. Yet training the developer relies on recognizable disruption cues from the ink, while improving the ink relies on reliable localization guidance from the developer. We address this mutual dependence through an alternating cooperative learning strategy that coordinates two complementary phases. In the edit-driven developer learning phase, we hold the current ink fixed and train the developer on actual edits of protected images. In the subsequent developer-guided ink learning phase, we hold the updated developer fixed and optimize the ink using a protected view and differentiable simulated views constructed by local ink erasure followed by image degradation. Repeatedly alternating these phases allows actual editing outcomes to guide ink learning through the developer, without backpropagating through the editors. Experiments on natural image and face datasets show that EditInk achieves higher performance than the compared passive and proactive baselines on both open and closed source editors.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image Forensics, Proactive Image Protection, Tamper Localization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_dRh4OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper embeds a protective signal before image release to help locate subsequent edits. EditInk supports black-box editors, and the experiments demonstrate localization across several editing settings.

Critical strengths

The experiments use actual editor outputs rather than region-composited images, making the evaluation closer to practical use.

Critical weaknesses

Comparisons on unseen API editors are missing. These would give a fuller picture of the method’s generalization.

Questions for authors
  1. Please evaluate EditInk on more unseen editors, including both local and API editors.
  2. Do the target masks for GPT/Grok outputs represent the intended semantic edits, or are they derived from before-and-after image differences? Please explain how the masks are constructed and how global appearance changes are handled.
  3. The paper has some grammatical issues that should be corrected. For example, “thus can supports any black-box editors” is ungrammatical; the sentence about the Face Data source and split in Section 5.1 needs rewriting; and the “Eq. equation” references should be fixed.
Reviewer_NTETOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

EditInk achieves active image protection and localization by alternately training ink and ink developer.The developer learns from actual edits, then guides ink updates through differentiable ink-erasure views. Experiments on natural images and faces evaluate the method with local and API editors.

Critical strengths

The two-stage training lets the ink and Ink Developer learn together from actual edits without backpropagating through the editor, offering a practical approach to proactive localization with black-box editors.

Critical weaknesses

The architectural choices for the Ink Developer need more experimental support.

Questions for authors
  1. Does the choice of DINO feature layer affect localization performance? Why use DINO as the feature extractor? Please provide experiments to support these choices.

  2. Why does local ink erasure simulate the changes caused by real edits? Please briefly explain the rationale.

  3. Some descriptions are inconsistent. Table 3 should be labeled EditInk*, and the text should state that the results are obtained after refinement. The setting names in Table 3 should also match those in the text.

Reviewer_1b8qOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EditInk learns a shared additive image signal and a manipulation-localization reader for forward-only editors. Training alternates between fitting a DINOv2-based reader to actual edited outputs and optimizing the signal through a frozen reader using differentiable local ink erasure and degradations. Deployment embeds the shared signal and predicts manipulated regions from the edited image alone. Evaluation covers natural and facial images, three open-source editors, two API editors, degradation robustness, and component ablations. The paper reports shared-signal API gains and separately evaluates a per-image PGD-refined variant.

Critical strengths
  1. The method accommodates forward-only editors. Fitting the localization head to actual editor outputs, then updating shared ink through a frozen head and differentiable erasure/degradation views, does not require editor gradients. Deployment uses shared ink and a single-image reader, which is a practical property for API editors.
  2. The reported shared-ink API results are promising. On HumanEdit, GPT F1 is .5111 versus WAM's .4071, and Grok F1 is .5539 versus .5120 (Table 1). Evaluation uses complete editor outputs rather than restoring unchanged backgrounds by compositing. These are positive reported results that remain relevant despite the unresolved ablation and metric issues.
Critical weaknesses
  1. The central ablation cannot be assigned confidently to the stated method. Section 5.4 says all ablations use shared ink without per-image PGD. Yet Table 3's Full EditInk IoUs, 65.99/68.21/86.53%, exactly match Table 1's EditInk* row, which uses 200 per-image PGD steps, rather than the shared row's 45.06/45.75/71.99%. Table 3's mean F1 is 82.99%, compared with 66.24% for the shared row and 82.98% for the starred row over those editors. This may be a labeling/reporting error, not proof that PGD was used, but it currently prevents reliable attribution of the component gains.
  2. The reported metrics require reconciliation. For the same per-image binary predictions, labels, sample weights, and compatible empty-mask handling, mean IoU <= mean F1 <= 2 mean IoU/(1+mean IoU). Table 2's Blank Canvas/GPT pair (.0578, .0107) violates the lower bound, and EditInk/Grok (.7246, .8767) exceeds the upper bound .84031. Table 1's OmniGuard/GPT pair (.2240, .3711) exceeds .36601. Printing precision cannot explain these differences. Different classes, thresholds, or sample sets could make the bounds inapplicable, but must be specified. The API difference-mask recipe also needs definition. These uncertainties affect the validity and comparability of the quantitative evidence, not just presentation.
Questions for authors
  1. Result-provenance clarification: Which configurations produced Tables 1/3 and Figures 5/6, including whether per-image PGD was used? Please reconcile the shared/starred rows and supply corrected ablation results or explain the matching values.
  2. Metric and label clarification: What exact predictions, thresholds, classes, averaging weights, and empty-mask conventions underlie each IoU/F1 pair? How are API ground-truth masks constructed and validated? Please explain the three conditional-bound violations or provide corrected, traceable aggregates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.