acceptodds
Under review as a conference paper at ICLR 2027

EditTrack: Detecting and Attributing AI-assisted Image Editing

Abstract

In this work, we formulate and study the problem of image-editing detection and attribution: given a base image and a suspicious image, detection seeks to determine whether the suspicious image was derived from the base image using an AI editing model, while attribution further identifies the specific editing model responsible. Existing methods for detecting and attributing AI-generated images are insufficient for this problem, as they focus on determining whether an image was AI-generated/edited rather than whether it was edited from a particular base image. To bridge this gap, we propose EditTrack, the first framework for this image-editing detection and attribution problem. Building on four key observations about the editing process, EditTrack introduces a novel re-editing strategy and leverages carefully designed similarity metrics to determine whether a suspicious image originates from a base image and, if so, by which model. We evaluate EditTrack on five state-of-the-art editing models across six datasets, demonstrating that it consistently achieves accurate detection and attribution, significantly outperforming five baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI Editing, Image Editing, Attribution

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

The authors study the problem of image-editing detection and attribution and evaluate on five models across six datasets. The author pioneerly addresses this task. However, some critical issues on fairness and reliability in experimental comparison exist and need clarifying.

Critical strengths
  1. A clear discrimination between editing detection given base-suspicious paired image, and attribution of editing model.
  2. A method of reference value simultaneously reporting detection and attribution, while attempting cross-model evaluation.
Critical weaknesses
  1. BLIP-2 is used to describe base and suspicious, which are then put together into proxy prompt, during which the target image semantics can be leaked into the re-editing pipeline.
  2. The reliability of ``dissimilarity/variety/stability'' is questionable given that only 50/100/2 positive/negative samples/editting models are validated on.
Questions for authors
  1. How does the detection/attribution accuracy change when using the original raw prompt rather than BLIP-2 proxy prompt?
  2. When totally unseen models are added into closed-set candidate models, is the system output Un-edited'' or Unseen editing models''?
Reviewer_o2BqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the problem of detecting and attributing AI-assisted image editing. It aims to determine whether a suspicious image was derived from a given original image through AI editing and, if so, to identify the editing model that was used. To this end, the authors propose EditTrack, a framework whose core idea is to re-edit both the original image and the suspicious image using candidate editing models, and then combine structural, semantic, and pixel-level similarity features with a multi-class classifier to perform detection and attribution. The authors conduct experiments on five image editing models and six datasets. The results show that EditTrack achieves high detection and attribution accuracy under the current experimental settings and outperforms the compared baselines.

Critical strengths
  1. Clear motivation.
  2. The problem studied in this paper has clear practical significance.
  3. The re-editing strategy is somewhat novel, and the experimental results are fairly convincing.
Critical weaknesses

I have the following questions regarding this paper:

  1. I have a major concern regarding the methodology. According to Section 4.2, the authors use BLIP-2 to generate captions for the original and suspicious images separately, combining them into a proxy editing prompt. However, independently generated captions may not accurately capture the editing differences between the two images. For instance, modifications to local colors, textures, or fine-grained attributes might not be reflected in the captions, leading to a significant discrepancy between the proxy prompt and the actual editing instructions. This is particularly critical because both Observation 1 and Observation 2 rely on the assumption that the proxy prompt is semantically close enough to the true edit. While the authors demonstrate the effectiveness of the re-editing strategy, they do not directly validate this assumption, nor do they compare different methods of prompt construction.
  2. Furthermore, the authors use the same proxy prompt (pe′p'_e) to re-edit both the original image (IbI_b) and the suspicious image (IsI_s). However, since the "original prompt" component within this proxy prompt describes (IbI_b), applying it directly to (IsI_s) creates a potential inconsistency between the input image and the description of the original state, which could affect the re-editing results.
  3. There are limitations to the experimental validation of the four observations. According to Section 4.2, the authors validate these observations using only two models, Step1X-Edit and EditAR, along with a dataset of 50 positive samples and 100 negative samples. Notably, according to the experimental setup described for Figure 2, Step1X-Edit uses the proxy prompt (pe′p'_e) to re-edit the original image, whereas EditAR uses the true editing prompt (pep_e). Consequently, differences between the two sets of results could stem from either the editing model or the prompt variation, making it difficult to isolate the model-specific differences highlighted in Observation 3. Additionally, the experiments do not directly show the similarity distributions for cases where an incorrect model re-edits a suspicious image or where the suspicious image in a negative pair is re-edited.
  4. I note that the negative samples used for training are drawn from MSCOCO, Artwork, and Inter4K, and none of the suspicious images in these sets have undergone AI-based editing. Although an "Unrelated" dataset was introduced during the testing phase, the AI-edited images within it were randomly paired with base images sampled from the negative test pairs; consequently, they may not adequately cover "hard negative" samples, namely cases where images are visually or semantically highly similar but lack a genuine editing relationship. As a result, the model might rely partly on artifacts of AI editing or differences in image content to make judgments, rather than accurately identifying the editing provenance relationship between the two images.
  5. The paper primarily compares EditTrack against VLM-based methods, watermarking techniques, and existing image attribution methods. Although some baselines take both the original and suspicious images as input, the paper lacks a simpler baseline that directly extracts structural and semantic similarity features from image pairs and trains a classifier without re-editing. Therefore, it remains unclear to what extent the improvement in detection performance comes from the re-editing mechanism itself, rather than the image-pair features or supervised classification. I encourage the authors to include such a baseline to better isolate the contribution of re-editing.
  6. The authors identify unknown editing models by setting a threshold for the "Non-edited" category, reporting a detection accuracy of 91% and an attribution accuracy of 83%. However, Table 3 in the appendix reveals that 21% of true negative samples were misclassified as edits by unknown models, and the correct rejection rate for negative samples was only 78.25%, which falls short of the 90% validation set accuracy targeted during threshold selection. This suggests that the current method may suffer from significant false-positive issues when identifying unknown models. Given the paper's emphasis on image provenance and forensic applications, I encourage the authors to further analyze how threshold selection affects false-positive rates and unknown model identification rates, and to report performance trade-offs across different thresholds, thereby providing a more comprehensive assessment of the method's reliability in open-set scenarios.
  7. I think the paper lacks a clear visual overview of the overall methodology. Although Section 4 describes the individual components of EditTrack, Figure 1 only illustrates the three categories of image similarity rather than presenting the complete framework.
Questions for authors

See the weakness.

Reviewer_6ZD9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper formulates image-editing detection and attribution: given a base image I_b and a suspicious image I_s, decide whether I_s was produced from I_b by an AI editing model, and if so, which of n candidate models produced it. The proposed method, EditTrack, rests on four claimed properties of editing models: robustness, stability, variety and dissimilarity. BLIP-2 first captions both images, and the captions are combined into a proxy editing prompt. Each candidate model then re-edits both I_b and I_s with this prompt, and six similarity metrics compare each re-edited image with I_s, covering structure (structural distance, pHash), semantics (CLIP, LPIPS) and pixel histograms (intersection, Bhattacharyya). The resulting 12n features are fed into a small MLP that classifies each pair into n+1 classes, one per candidate model plus "not edited".

Critical strengths
  1. The paper tackles a problem that matters in practice. Asking whether a suspicious image was edited from a specific base image, and by which model, is a question artists and forensic analysts actually face, and standard single-image detectors can't answer it. The method is also appealingly simple: it treats the editing models as black boxes, needs neither their weights nor the original editing prompt, and relies on a lightweight classifier on top of interpretable similarity features.

  2. The idea is easy to follow. The re-editing intuition is clearly explained, and Figure 2 gives a reasonable sanity check of the underlying assumptions. EditTrack clearly outperforms the adapted baselines on both detection and attribution, and I appreciated the ablations on feature groups and metrics as well as the unseen-model experiment.

Critical weaknesses
  1. My first concern is that the evaluation does not show that re-editing is necessary. Positive pairs preserve the base image's layout and content, while negatives are simply different images, so computing the same six metrics directly on (I_b, I_s) might already separate them. This baseline is missing, and copy-detection methods such as SSCD are not compared. Besides, the negatives are also too easy: three of the four test sets contain no AI edits, and hard negatives, such as AI edits of a visually similar but different base image, are absent. With only 120 training and 100 test base images and no variance reported, it is unclear how much of the performance comes from the proposed mechanism.

  2. Second, I am also worried about that the test edits are favourable to the method. They use the same prompt template as EditTrack's proxy prompt ("Do the image editing task; origin prompt: …, editing prompt: …"; Figure 5), and most are whole-object replacements written as caption substitutions. Free-form instructions, local or subtle edits, style edits, multi-step or cross-model edits, and evasive users are not evaluated. Since EditTrack relies on the proxy prompt approximating the true one, this is a key limitation for real-world use.

Questions for authors
  1. What are detection and attribution accuracy when the same six metrics are computed directly between I_b and I_s, without re-editing?

  2. How does EditTrack perform on hard negatives, where I_s is an AI edit of a different but visually similar base image?

  3. Why are Tables 5 and 6 identical, and why does Table 6 disagree with Table 2 for QWen-FT?

  4. Does attribution accuracy hold when all suspicious images are resized to a common resolution and JPEG-compressed before evaluation?

  5. How does performance change when test edits use free-form instruction prompts written independently of the proxy template?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.