acceptodds
Under review as a conference paper at ICLR 2027

Reading the Visual Clues: A Dataset and Contextual Detector for Document Tampering

Abstract

Document tampering extends beyond edited text to figures, formulas, and tables, motivating a common evaluation of editing operations across page elements. We introduce MEDTD, a benchmark for localizing operation-involved regions and recognizing editing operations from a single full-page image. Constructed from DocLayNet, MEDTD contains 46,290 programmatically edited pages, 25,145 untampered counterparts, and 69,839 region annotations covering seven element groups and four operations. Its annotations include unchanged duplication sources and vacated relocation sources, rather than modified pixels alone. We further introduce ClueDet, a DINO-based detector that refines detection queries through similarity-weighted comparisons of predicted regions. Under matched training settings, ClueDet achieves 34.11 operation AP, compared with 33.02 for DINO and 32.56 for a parameter-matched uniform-context variant. Improvements are concentrated in textual elements and splicing rather than shared uniformly across element groups. Evaluation on untampered pages and external datasets characterizes false-positive behavior and limits to transfer. MEDTD provides a benchmark for studying operation-aware document tampering across heterogeneous page elements.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Document tampering detection, Image forensics, Region-context modeling, Document image datasets

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nXdgOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
1
Read review
Paper summary

This work constructs the MEDTD dataset for document tampering detection, which incorporates not only textual content but also visual elements commonly embedded in documents, including images, formulas, and tables. Furthermore, the ClueDet model is proposed based on DINO. It performs feature aggregation using visual similarity across different regions within a page to optimize detection queries. Experiments verify the effectiveness of this context-aware contrastive paradigm.

Critical strengths
  1. This work builds the multi-element document tampering dataset MEDTD, covering diverse content such as texts, figures, tables, and formulas, with labels annotated for four categories of tampering operations.
  2. The ClueDet detector is proposed to aggregate contextual features via intra-page regional similarity, validating the benefits of page-level contextual information for document tampering recognition.
Critical weaknesses
  1. The motivation for constructing the dataset is insufficiently highlighted. Although this dataset additionally introduces editing and tampering of graphics and tables, it bears no fundamental distinction from existing document manipulation/forgery datasets. In particular, as shown in Table 1, the proposed dataset exhibits no distinctive advantages in terms of the counts of original/tampered images, document categories, or manipulation types (CE, CD, CR, CDS). Moreover, the workload invested is not substantial. It is suggested to explore broader forms of adversarial editing, such as reprinting and regional AI generation, to enrich the dataset coverage. Therefore, I argue that the dataset construction is a weak core contribution.

  2. The proposed ClueDet (presented in Section 3.2 and Figure 2) lacks sufficient novelty. It essentially assembles existing components (DINO, FFN) without dedicated improvements tailored to document priors, especially forgery clues including CE, CD, CR, and CDS.

Given the two major drawbacks outlined above, I cannot assign a positive score to this manuscript.

Questions for authors

Please refer to the weaknesses.

3 comments

We thank Reviewer nXdg for the feedback and comments. We respectfully disagree that MEDTD bears no fundamental distinction from existing datasets, and we would like to clarify the distinction with the evidence below.

[W1] Motivation and distinction of the dataset

The motivation for constructing the dataset is insufficiently highlighted. Although this dataset additionally introduces editing and tampering of graphics and tables, it bears no fundamental distinction from existing document manipulation/forgery datasets. In particular, as shown in Table 1, the proposed dataset exhibits no distinctive advantages in terms of the counts of original/tampered images, document categories, or manipulation types (CE, CD, CR, CDS). Moreover, the workload invested is not substantial. It is suggested to explore broader forms of adversarial editing, such as reprinting and regional AI generation, to enrich the dataset coverage. Therefore, I argue that the dataset construction is a weak core contribution.

(1) Motivation. A mask of modified pixels cannot represent the unchanged source of a duplication, and it cannot tell whether a blank region was erased or its content was moved to another place. Therefore, MEDTD annotates both regions of duplication and relocation events and evaluates the operation assigned to each box. Source and target localization has a precedent in Forensim (Sec. 2), and MEDTD brings such targets to full document pages.

(2) Comparison in Table 1. The following table compares MEDTD with the seven previous datasets on the columns of Table 1.

Column of Table 1Seven previous datasetsMEDTD
Edited elementsText only (numeric text for AIForge-Doc)Words, lines, paragraphs, pictures, formulas, tables and table-cell content
Source documentsNot reported1,995 recipient documents
Tampered images4,061 to 793,73146,290, more than four of the seven

As shown in the table, MEDTD is the only dataset whose edits go beyond text and the only one that reports document identities, which allows us to keep all pages of a recipient document in one split. All 28 combinations of operation and element are present in every split. Besides, its pages come from six source domains, and the domain matters: in all full-data DINO-family runs reported, the operation AP is highest on laws and regulations (41.83–46.78) and scientific articles (38.38–41.75) and stays at 19.83–31.36 on financial reports, manuals, patents and government tenders. We agree that the four operations are standard. The distinction lies in which content is edited and what is annotated, i.e., both regions of duplication and relocation events with their operation, together with 25,145 untampered pages for false-positive evaluation.

(3) Evidence for the distinction. If MEDTD had no distinction from existing datasets, released forgery localizers should transfer to it. The following table lists the region AP on the tampered test pages and the false-positive rate on the untampered test pages (Table 4).

ModelRegion APFalse-positive rate
ADCD-Net0.9951.12%
Co-Transformers0.0269.48%
TruFor0.09100.00%
PIXAR0.61100.00%
DCL-Net13.6222.21%
DINO34.178.79%

As shown in the table, the released models obtain below 1 region AP and flag more than half of the untampered pages. Even after training on MEDTD, DCL-Net as the best of seven forensic localizers reaches only 13.62 region AP, compared with 34.17 for DINO. The element type also matters, e.g., in all runs the AP is 42.17–49.71 for duplicated pictures but only 0.11–0.50 for duplicated words, which a text-only dataset cannot show. Finally, in a post-hoc analysis of the nine runs, 94.0–95.2% of the vacated relocation sources are overlapped by an exported box, but only 43.9–53.5% of them receive the relocation label from the highest-scoring box, compared with 87.2–92.9% correct labels for the three other region types we examined.

(4) Workload. We inspected every one of the 46,290 edited pages, corrected errors and reinspected the pages (Sec. 3.1.3), which results in 69,839 region annotations. Table 4 evaluates twenty methods from four families under one protocol, and seven of them are forensic localizers trained on MEDTD. The data volume is also important for training: with fixed optimizer updates, training on the full partition instead of a fixed half improves the operation AP by 12.16 points for DINO and 1.86 for Faster R-CNN, and reduces the false-positive rate of DINO from 28.81% to 8.79% (Tables 5 and 10).

(5) Reprinting and regional AI generation. Thank you for these suggestions. Capture conditions and generated text are the focus of other datasets in Table 1, i.e., FSTS (scans and captures), DanceText (AI-generated and edited text) and AIForge-Doc (generative field edits), which all edit text only, and MEDTD is complementary to them. Following your reprinting suggestion, we nevertheless evaluated a synthetic print–scan degradation (half resolution, blur, gain and noise) on the full test set, for the forensic localizers of Table 4 (region AP).

ModelOriginalPrint–scanRetained
DINO34.174.9214.4%
Uniform control34.005.3815.8%
ClueDet35.133.8611.0%
RTMDet-M20.213.1415.5%
Cascade R-CNN19.611.668.5%
RetinaNet18.070.643.5%
Faster R-CNN17.240.583.3%
FRD-Net13.169.7173.8%
CAFTB-Net†9.908.7988.8%
DCL-Net13.627.0051.4%
TIFDM†9.133.9042.8%
PSCC-Net3.201.8557.8%
CAT-Net v2†3.850.000.0%

The results show that all seven box detectors keep only 3–16% of their region AP, while FRD-Net, CAFTB-Net and DCL-Net keep 51–89% and overtake them. Therefore, MEDTD can already be used to study such degradations, and the ranking of the evaluated models changes under them. Physical reprinting and regional AI generation are not part of the current release, and we will name both as limitations.

[W2] What is tailored to forgery clues in ClueDet

The proposed ClueDet (presented in Section 3.2 and Figure 2) lacks sufficient novelty. It essentially assembles existing components (DINO, FFN) without dedicated improvements tailored to document priors, especially forgery clues including CE, CD, CR, and CDS.

We would like to clarify that the FFN is only the final projection of the added module. Before it, ClueDet pools the features inside each predicted box, looks for similar regions on the same page, and passes the difference between a region and their weighted mean to the decoder query (Sec. 3.2, Fig. 2). This design follows from the operations you mention. A duplication leaves a second copy on the page, a relocation leaves a vacated source and a displaced copy, and a splicing inserts content from another page. In each case the evidence lies in how a region relates to other regions, which is what the module compares. The training keeps the classification and box losses of DINO and uses no contrastive loss.

The effect of this comparison can be tested with controls of matched size. In addition to the uniform-aggregation control in the submission, we trained a query-only MLP without any region input and a local-feature refinement that uses the region feature but makes no comparison. All models use the same training setting at the submitted seed.

ModelOperation APClueDet minus model
DINO33.02+1.09
Query-only MLP33.51+0.60
Local-feature refinement32.38+1.73
Uniform aggregation32.56+1.55
ClueDet34.11

The results show that, with the same regions, similarity weighting is better than uniform averaging and better than using the local region feature without any comparison. However, the query-only MLP is not separated from ClueDet, so we do not attribute the gain over DINO to the regional comparison itself.

We also trained DINO and ClueDet with two additional seeds. The DINO and ClueDet columns are operation AP.

SeedDINOClueDetOperation AP differenceRegion AP difference
033.0234.11+1.09+0.96
132.1732.65+0.48+0.62
233.2235.22+2.00+1.75

We appreciate your comments and hope this clarification of the distinctions and contributions addresses your concerns. If you know an existing dataset with edited pictures, formulas, and tables, or with annotated duplication and relocation sources on full document pages, we would be glad to compare it with ours. We would be grateful if you would reconsider the dataset contribution and the paper's soundness based on what MEDTD annotates, what it covers, and what it measures.

Reviewer_X1j4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript presents a dataset and a contextual detector for document tampering detection and localization. It introduces MEDTD, which covers multiple document elements and tampering operations, and ClueDet, a detector built on DINO. ClueDet aggregates features from visually similar regions and uses the resulting contextual features and their differences from local features to refine detection queries.

Critical strengths
  1. The proposed MEDTD dataset is valuable for studying document tampering across text and non-text elements within a common evaluation framework.
  2. ClueDet has a straightforward design, using similarity-weighted regional comparisons to refine DINO’s detection queries.
Critical weaknesses

1. Limited methodological novelty.

ClueDet combines region sampling, cosine-similarity aggregation, and residual query refinement within DINO. While the design is intuitive, the paper does not clearly establish a methodological advance over existing attention-based context modeling.

2. Insufficient experimental validation.

The uniform-aggregation control isolates similarity weighting but does not establish the benefit of regional comparison over a parameter-matched query MLP, local-feature refinement, or standard regional attention. The gain is modest and concentrated in splicing; multiple training seeds and direct tests of source–destination correspondence are needed to support the proposed mechanism.

3. Limited evidence of robustness under real dissemination.

MEDTD is generated by editing digital PNG pages and lacks controlled tests of screenshot resampling, recompression, or camera capture. The external-dataset evaluation is useful, but changes in tampering operations, annotations, and imaging conditions are confounded. It therefore cannot establish whether the detector’s gains survive common dissemination processes.

Questions for authors
  1. What does ClueDet’s regional comparison capture beyond existing attention-based context modeling?
  2. Does the gain persist across training seeds against parameter-matched query MLP, local-feature refinement, and standard regional attention baselines?
  3. Does ClueDet retain its advantage over DINO under controlled screenshot resampling, recompression, and camera capture?
3 comments

We thank Reviewer X1j4 for the valuable feedback, for recognizing the value of MEDTD for studying tampering across text and non-text elements, and for the specific suggestions on controls, training seeds and dissemination tests. We ran the requested experiments and report all results.

[W1 & Q1] What regional comparison captures beyond attention

Limited methodological novelty. ClueDet combines region sampling, cosine-similarity aggregation, and residual query refinement within DINO. While the design is intuitive, the paper does not clearly establish a methodological advance over existing attention-based context modeling.

Q1: What does ClueDet’s regional comparison capture beyond existing attention-based context modeling?

In DINO, the self-attention exchanges query representations and the deformable cross-attention samples encoder features, but neither of them explicitly compares what two predicted boxes contain. ClueDet adds this comparison. It pools the image features inside each predicted box, builds a reference from the other boxes on the same page (weighted by cosine similarity, excluding the box itself and boxes that overlap it at IoU ≥ 0.5), and gives the query its own pooled feature together with the difference from this reference. The motivation is that documents repeat units, e.g., lines and table cells, so a region can be compared with similar regions on the same page.

The effect of the similarity weighting can be tested directly, because the uniform control uses the same regions and the same number of parameters. Similarity weighting achieves 34.11 operation AP, compared with 32.56 for uniform weights. This advantage is also kept in the dissemination tests reported below.

[W2 & Q2] Training seeds, matched controls and correspondence

Insufficient experimental validation. The uniform-aggregation control isolates similarity weighting but does not establish the benefit of regional comparison over a parameter-matched query MLP, local-feature refinement, or standard regional attention. The gain is modest and concentrated in splicing; multiple training seeds and direct tests of source–destination correspondence are needed to support the proposed mechanism.

Q2: Does the gain persist across training seeds against parameter-matched query MLP, local-feature refinement, and standard regional attention baselines?

Thank you very much for these suggestions. We trained DINO and ClueDet with two further seeds, using the same 32,403 training pages, 12 epochs, and the final checkpoint. The DINO and ClueDet columns are operation AP, and each interval uses 2,000 paired resamples of the 399 test documents.

SeedDINOClueDetOperation AP differenceRegion AP difference
033.0234.11+1.09+0.96
132.1732.65+0.48+0.62
233.2235.22+2.00+1.75

As shown in the table, the difference is positive.

We also trained the two requested controls. All refinements use the same insertion points, and their parameter counts are within 0.1% of ClueDet.

RefinementOperation APClueDet minus control
Query-only MLP33.51+0.60
Query and local region feature32.38+1.73
Uniform regional aggregation32.56+1.55
ClueDet34.11

The results show that ClueDet is separated from the two controls that use region features without the similarity-weighted comparison, but not from the query-only MLP. This control also recovers much of the splicing gain (45.74 AP, compared with 43.21 for DINO and 46.35 for ClueDet) and reaches the line-level values of ClueDet (7.85 and 53.09 AP for duplicated and spliced lines, compared with 7.93 and 53.24). Therefore, we do not attribute the gain over DINO to the regional comparison. For completeness, relative to DINO the query-only MLP is +0.49, and the local-feature control is −0.64.

Regarding the source–destination correspondence, ClueDet predicts labelled boxes and no link between them, so we tested the correspondence on the outputs. For the 2,409 relocation events of the test set, we measured how often both regions are recovered, i.e., matched by a relocation box with a score of at least 0.5 at IoU ≥ 0.5, in the full-data DINO-family runs reported. The destination is recovered for 7.6–13.7% of all events, but for 40.0–53.1% of the events whose vacated source is recovered, i.e., the two regions of an event are found together.

[W3 & Q3] Robustness under dissemination

Limited evidence of robustness under real dissemination. MEDTD is generated by editing digital PNG pages and lacks controlled tests of screenshot resampling, recompression, or camera capture. The external-dataset evaluation is useful, but changes in tampering operations, annotations, and imaging conditions are confounded. It therefore cannot establish whether the detector’s gains survive common dissemination processes.

Q3: Does ClueDet retain its advantage over DINO under controlled screenshot resampling, recompression, and camera capture?

We agree with you that the external-dataset evaluation changes images, operations and annotations at the same time. To separate these factors, we kept the 9,258 edited and 4,849 untampered test pages, their labels and only transformed the images.

TestDINOClueDetClueDet minus DINOClueDet minus uniform control
Original pages33.0234.11+1.09+1.55
JPEG, quality 9531.1232.81+1.68+1.84
JPEG, quality 8527.9329.58+1.65+2.32
JPEG, quality 7525.6826.82+1.13+1.95
Resampling to 0.75× and back31.6531.78+0.13+1.22
Resampling to 0.5× and back28.3628.63+0.27+1.43
Synthetic print–scan (half resolution, blur, gain, noise)3.131.85−1.28−2.29

The results show that the advantage over DINO survives recompression, is not established after resampling, and is lost under the print–scan degradation. The advantage over the uniform control is kept in the five compression and resampling tests and is also lost under the print–scan degradation. Both models keep 78–96% of their operation AP in the five compression and resampling tests.

The print–scan test also gives a result about the benchmark. We applied it to the other box detectors and to six MEDTD-trained forensic localizers of Table 4.

ModelOriginalPrint–scanRetained
DINO34.174.9214.4%
Uniform control34.005.3815.8%
ClueDet35.133.8611.0%
RTMDet-M20.213.1415.5%
Cascade R-CNN19.611.668.5%
RetinaNet18.070.643.5%
Faster R-CNN17.240.583.3%
FRD-Net13.169.7173.8%
CAFTB-Net†9.908.7988.8%
DCL-Net13.627.0051.4%
TIFDM†9.133.9042.8%
PSCC-Net3.201.8557.8%
CAT-Net v23.850.000.0%

As shown in the table, all seven box detectors keep only 3–16% of their region AP, while FRD-Net, CAFTB-Net and DCL-Net keep 51–89% and end up above every box detector.

We hope these experiments can answer your three questions. Thank you again for your constructive suggestions.

Reviewer_ksa7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces MEDTD, a benchmark for operation-aware document tampering detection from a single full-page image. Built from DocLayNet, MEDTD contains 46,290 programmatically edited pages, 25,145 untampered pages, and 69,839 region annotations covering seven element groups and four editing operations. The authors also propose ClueDet, a DINO-based detector that refines detection queries through similarity-weighted comparisons of predicted regions. Under matched training settings, ClueDet reports 34.11 operation AP versus 33.02 for DINO and 32.56 for a parameter-matched uniform-context variant, with gains concentrated in textual elements and splicing. The paper further analyzes false positives on untampered pages and transfer limits on external datasets.

Critical strengths
  1. The benchmark broadens document tampering beyond text editing to figures, formulas, and tables, enabling a more unified evaluation of operation-aware tampering across heterogeneous page elements.

  2. The annotation design is thoughtful: unchanged duplication sources and vacated relocation sources provide more informative supervision than modified pixels alone.

  3. The benchmark includes untampered pages, supporting false-positive analysis, and the paper explicitly studies transfer limits rather than claiming universal generalization.

  4. ClueDet is conceptually simple and integrates naturally with a DINO-style detector, with matched and parameter-matched comparisons.

Critical weaknesses
  1. Synthetic edits may introduce detectable artifacts and do not establish real-world validity. MEDTD is built by programmatic editing from DocLayNet; detectors may exploit rendering, boundary, or layout artifacts rather than genuine tampering cues. Real tampered documents or stronger artifact-controlled edits are needed.

  2. The operation taxonomy and annotation semantics need clearer formalization. The definitions of the seven element groups, four operations, duplication/relocation sources, and ambiguous or overlapping cases are not sufficiently specified, nor is annotation consistency or quality control.

  3. The empirical gain is modest and narrowly localized. ClueDet improves operation AP from 33.02 to 34.11 over DINO, with gains concentrated in textual elements and splicing; per-operation, per-element, significance, and localization-quality analyses are needed to assess robustness.

  4. The methodological novelty of ClueDet is limited and insufficiently isolated. Similarity-weighted query refinement is a reasonable but incremental extension of DINO; stronger ablations, controls for synthetic artifacts, and comparisons with tampering-specific detectors are needed.

Questions for authors

Please refer to the weaknesses.

4 comments

We thank Reviewer ksa7 for the valuable feedback, and for recognizing that MEDTD broadens document tampering beyond text and that its annotation of unchanged duplication sources and vacated relocation sources is more informative than modified pixels alone. We address your concerns below.

[W1] Synthetic artifacts and real-world validity

Synthetic edits may introduce detectable artifacts and do not establish real-world validity. MEDTD is built by programmatic editing from DocLayNet; detectors may exploit rendering, boundary, or layout artifacts rather than genuine tampering cues. Real tampered documents or stronger artifact-controlled edits are needed.

Thank you for raising this point. We chose programmatic editing because the regions that MEDTD is about cannot be annotated on forgeries that are collected afterwards. For example, an unchanged duplication source contains no modified pixel, and whether a blank region was erased or its content was moved elsewhere is known only from the edit record. To ensure the quality, we inspected all 46,290 generated pages, corrected errors and inspected them again. Besides, we examined in three ways whether the detectors only follow traces of the editing routine.

(1) Targets without any trace. The 2,301 unchanged duplication sources of the test set contain no modified pixel, so no rendering or boundary artifact can reveal them, and they are evaluated like every other target. In the full-data DINO-family runs reported with various seeds and controls, only 0.26–4.09% of them are recovered at score 0.5, and duplication is the hardest operation in every run (10.91–13.24 AP). Therefore, MEDTD cannot be solved by detecting artifacts, and it measures how far a model is from the targets that need more than a local trace.

(2) Same routine, different labels. We compared targets which are produced by the same routine but have different labels. The label of a target is taken from the highest-scoring predicted box that overlaps it at IoU ≥ 0.5.

RoutineTargetCorrect label, DINO / ClueDet
PasteInserted duplicate (CD)88.1% / 89.9%
PasteRelocation destination (CR)92.1% / 92.9%
FillErased region (CE)91.3% / 90.0%
FillVacated relocation source (CR)45.2% / 47.8%

As shown in the table, the two pasted targets are labelled correctly in about 90% of the cases although they come from the same paste routine, so the predicted operation does not simply follow the local routine. For the two filled targets the separation is only partial, since about half of the vacated sources are predicted as erasure.

(3) Same edits, changed pixels. We kept the edits, the labels and the seed fixed and only transformed the images of the 9,258 edited and 4,849 untampered test pages. The values are operation AP, with the retained share of the original AP in brackets. JPEG compression is one encode–decode cycle.

TestDINOClueDet
Original pages33.0234.11
JPEG, quality 9531.12 (94.3%)32.81 (96.2%)
JPEG, quality 8527.93 (84.6%)29.58 (86.7%)
JPEG, quality 7525.68 (77.8%)26.82 (78.6%)
Resampling to 0.75× and back31.65 (95.8%)31.78 (93.2%)
Resampling to 0.5× and back28.36 (85.9%)28.63 (83.9%)
Synthetic print–scan (half resolution, blur, gain, noise)3.13 (9.5%)1.85 (5.4%)

The results show that both models keep 78–96% of their operation AP when every page is recompressed or resampled, so the predictions do not depend only on fine pixel-level traces of the editing routine. In contrast, the synthetic print–scan degradation removes most of the AP. In this test all detectors of Table 4 keep only 3–16% of their region AP, while MEDTD-trained forensic localizers (DCL-Net, CAFTB-Net and FRD-Net) keep 51–89% and end up above every box detector, at 7.00–9.71 compared with at most 5.38 region AP. This degradation is not calibrated to a real printer or scanner.

[W2] Operation taxonomy and annotation semantics

The operation taxonomy and annotation semantics need clearer formalization. The definitions of the seven element groups, four operations, duplication/relocation sources, and ambiguous or overlapping cases are not sufficiently specified, nor is annotation consistency or quality control.

Thank you for this suggestion. We summarize the definitions below and will add them as a table to Sec. 3.1.

(1) Element groups. The seven element groups are words, lines, paragraphs, pictures, formulas, tables and table-cell content. For nested elements, the unit that was actually edited determines the label, e.g., editing one line inside a paragraph is a line event, and editing the text of a table cell is different from editing the whole table. All 28 combinations of operation and element group are present in the training, validation and test splits (Table 2 and Appendix A).

(2) Operations and annotated regions. Each edited page contains exactly one editing event, and the operation defines which regions are annotated.

OperationAnnotated regionsBoxes in the test set
Content erasure (CE)The region whose content was removed3,105
Content duplication (CD)The unchanged source and the inserted copy4,602
Content relocation (CR)The vacated source and the destination4,818
Cross-document splicing (CDS)The receiving region only, because the donor is on another page1,443

(3) Sources and ambiguous cases. For CD and CR, both boxes share the operation label and are evaluated independently, and no copying direction has to be predicted. Ambiguous cases are resolved by the edit record, e.g., a filled region is labelled as CR if its content was moved to another position on the page, and as CE if the content was removed.

(4) Consistency and quality control. The operation labels come from the edit record. We inspected all 46,290 edited pages, corrected errors and reinspected them, removed four duplicate source groups with their eight derivatives, and verified that the final boxes contain 99.715% of the changed pixels. More details of the procedure can be found in Appendix A.2.

[W3] Size and distribution of the gain

The empirical gain is modest and narrowly localized. ClueDet improves operation AP from 33.02 to 34.11 over DINO, with gains concentrated in textual elements and splicing; per-operation, per-element, significance, and localization-quality analyses are needed to assess robustness.

Thank you for asking for these analyses. Most of them are in the submission (Tables 4, 10, 11 and 12), and we summarize them here together with the new results. All values in the first table are for the submitted seed.

AnalysisDINOClueDet
Operation AP33.0234.11
Region AP34.1735.13
Erasure / duplication / relocation AP40.93 / 12.74 / 35.1941.02 / 13.24 / 35.83
Splicing AP43.2146.35
Region AP on text / visual elements30.76 / 39.2233.01 / 38.83
Operation AP at IoU 0.7536.0036.76

Overall, the gain is +1.09 operation AP and +0.96 region AP (0.43 to 1.51). It appears in all four operations and is largest for splicing, and it is kept at the stricter IoU threshold of 0.75. On the element level the gain comes from text elements, e.g., line-splicing AP rises from 46.13 to 53.24, while the AP on visual elements is slightly lower. At fixed operating points, ClueDet recalls more targets than DINO at thresholds calibrated to a 1% validation false-positive rate (+0.93 points) and at score 0.5 (+1.99 points), both with intervals above zero, and the two do not differ at 5%. The per-operation and per-element values are descriptive and have no intervals.

Regarding robustness, we trained DINO and ClueDet with seeds, using the same 32,403 training pages, 12 epochs and the final checkpoint. The DINO and ClueDet columns are operation AP, and each interval uses 2,000 paired resamples of the 399 test documents.

SeedDINOClueDetOperation AP differenceRegion AP difference
033.0234.11+1.09+0.96
132.1732.65+0.48+0.62
233.2235.22+2.00+1.75

In the revised paper, we will report all three seeds and state the result as a controlled comparison with the submitted seed.

[W4] Novelty, ablations and tampering-specific baselines

The methodological novelty of ClueDet is limited and insufficiently isolated. Similarity-weighted query refinement is a reasonable but incremental extension of DINO; stronger ablations, controls for synthetic artifacts, and comparisons with tampering-specific detectors are needed.

We agree that ClueDet is built from standard components. Our aim was not a new primitive but a document-specific use of region content: each predicted region is compared with the content of the other predicted boxes on the same page, because duplication, relocation, and splicing leave their evidence in the relation between regions rather than inside a single region.

Regarding stronger ablations, in addition to the uniform-aggregation control in the submission, we trained two new parameter-matched controls, i.e., a query-only MLP without any region input and a local-feature refinement that uses the region feature but makes no comparison. All models use the same training setting at seed 0.

ModelOperation APClueDet minus model
DINO33.02+1.09
Query-only MLP33.51+0.60
Local-feature refinement32.38+1.73
Uniform aggregation32.56+1.55
ClueDet34.11

The results show that the similarity-weighted comparison is better than using the local feature only or averaging the same regions. However, the difference to the query-only MLP has an interval that includes zero, and this control also reaches the element-level values of ClueDet, e.g., 53.09 AP for spliced lines. Therefore, we cannot attribute the gain over DINO to the regional comparison itself, and we will state this in the revised paper.

Regarding tampering-specific detectors, the following table lists their region AP and their false-positive rate on the untampered test pages.

MethodRegion APFalse-positive rate
DCL-Net13.6222.21%
FRD-Net13.1690.55%
CAFTB-Net†9.9020.71%
TIFDM†9.1312.11%
FLAME6.4033.06%
CAT-Net v23.8512.44%
PSCC-Net3.2086.53%
ADCD-Net0.9951.12%
TruFor0.09100.00%
Co-Transformers0.0269.48%
DINO34.178.79%
ClueDet35.139.53%

As shown in the table, the best forensic localizer trained on MEDTD reaches 13.62 region AP, compared with 34.17 for DINO, and the released models stay below 1. Besides, fusing existing outputs does not reach the result of ClueDet either: aligning DINO boxes with DCL-Net proposals gives 33.13 operation AP, and weighted box fusion of DINO and Faster R-CNN gives 29.38.

Finally, we would like to note that the findings of the benchmark do not depend on these method results. In all runs reported in this response, duplication is the hardest operation, and only 43.9–53.5% of the covered vacated relocation sources receive the relocation label. We hope that the formalized annotation rules, the three analyses on synthetic artifacts, and the complete report of the method results can address your concerns, and we would be grateful if you could reconsider your rating. Thank you again for your thoughtful review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.