acceptodds
Under review as a conference paper at ICLR 2027

Detect the Forgery, Not the Visual Disturbances: MLLM-Guided Representation and Selective Modulation for ID Document Images

Abstract

ID document image tampering detection and localization is an important problem in multimedia security, where detection determines whether an image has been tampered with, while localization identifies the tampered regions. Existing methods primarily focus on clean ID document image conditions. However, Visual Disturbances (VD), such as glare, mosaics, or stains, may occur in ID document images; their low-level features are highly similar to tampered artifacts, potentially leading to VD Induced Forensic Misattribution (VDFM). To address this problem, we propose CLEAR-ID, a new approach for ID document VD scenarios that integrates forensic experts with Multimodal Large Language Model (MLLM). CLEAR-ID uses prior information introduced by an MLLM to adaptively represent forensic expert evidence as VD-oriented or tampered-oriented cues, and projects the resulting features into a low-rank residual space to constrain the modulation of forensic expert evidence. To fill the data gap in this scenario, we further introduce DISTURBID-180k. It covers 42 types of VD and provides corresponding severity and localization mask annotations, enabling systematic analysis of the impact of VD on tampered detection and localization. Experiments demonstrate that VD-aware evidence representation and constrained VD-guided modulation effectively mitigate VDFM. Across diverse VD, CLEAR-ID improves both image-level detection and pixel-level localization over the state-of-the-art (SOTA) forensic expert baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Forensics, ID Document Image Tampered Detection and Localization, Multimodal Large Language Models, Visual Disturbances

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MR9iOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies visual-disturbance-induced forensic misattribution (VDFM) in ID document images: non-malicious disturbances may be mistaken for tampering, while suppressing disturbance responses may cause genuine tampering to be missed. The authors introduce DISTURBID-180k, a dataset covering 42 types of visual disturbances, and propose CLEAR-ID. The method uses semantic and spatial priors from an MLLM-based disturbance localizer to form disturbance-oriented and tampering-oriented forensic representations. Low-rank selective modulation then constrains prior-conditioned updates to the tampering features.

Critical strengths
  1. The problem is practically relevant. The paper addresses both false positives caused by common disturbances such as glare and watermarks, and missed tampering caused by excessive disturbance suppression. These concerns reflect practical needs and warrant further study.
  2. The dataset and evaluation cover a broad range of conditions. The dataset provides separate annotations for disturbance and tampering regions and includes tampering in text, portrait, and signature fields. Evaluation across different editing systems, unseen disturbances, physical capture disturbances, and an external data domain helps assess performance under varied conditions.
Critical weaknesses
  1. Comparisons of false positives and false negatives with the closest prior work are missing. The paper does not report disturbance-specific FPR and tampering-specific FNR for closely related methods such as DCLNet. It therefore remains unclear whether CLEAR-ID is better at avoiding false alarms on benign disturbances. See Question 1.
  2. The contribution of MLLM semantics is not isolated. In Table 2, the full prior and the ground-truth disturbance masks differ in both the information they provide and their representation. The performance difference cannot be attributed specifically to the MLLM’s semantic capabilities.
  3. The functional roles of the two feature branches are not directly validated. The experiments mainly report final detection performance, without separately evaluating the predictive capabilities of the two branches or their roles in regions where disturbances overlap with tampering. The additional benefit of the provenance-aware design over standard multi-task learning remains unclear.
  4. Qualitative comparisons are limited. Figure 8 omits important baselines such as DCLNet and the MAER-only configuration, making it difficult to assess the effects of individual modules visually. Representative examples in the main paper, with enlarged crops and explicit annotations of false positives and false negatives, would help show the method’s performance and limitations in challenging regions.
Questions for authors
  1. The aggregate metrics in Table 1 do not separately report disturbance-related false positives and missed tampering, and F1 does not distinguish their contributions to changes in performance. The TTDMamba module ablations in Table 3 do not replace a direct comparison with closely related methods such as DCLNet. Reporting disturbance-specific FPR and tampering-specific FNR under the same metric definitions and evaluation protocol would provide more direct evidence of CLEAR-ID’s advantage on the central problem.
  2. In Table 2, the full prior and the ground-truth disturbance masks differ in both information content and representation. This comparison supports the effectiveness of the full prior configuration, but does not attribute the improvement specifically to the MLLM’s semantic understanding. Further controlled comparisons are needed to distinguish the contributions of semantic information and spatial priors while accounting for differences in training data, supervision, and prior representation.
  3. Different supervision tasks provide reasonable training objectives for the disturbance and tampering branches. However, improved final detection performance alone does not establish the additional value of the provenance-aware representation design. A comparison with a standard multi-task model with comparable parameter counts, supervision tasks, and training settings would help distinguish the benefits of the proposed design from those of additional auxiliary supervision.
  4. Although the authors allow the two representations to share information, their intended functional roles still require direct validation. Can both FdF_d and FtF_t predict visual disturbances, and how well does each support tampering localization? What roles do the two branches play in regions where disturbances and tampering overlap, and where they do not?
  5. The paper shows reduced false positives relative to the clean-only baseline and some ablation configurations, and LSM reduces the disturbance-related false positives of MAER. However, in Table 3, the full model has a higher UVD FPR than VD-aware TTDMamba (0.35 versus 0.30), but a lower FNR (31.68 versus 38.47). This result reflects a trade-off between false positives and false negatives. How do the authors explain this behavior? At the same tampering recall, does CLEAR-ID still achieve fewer disturbance-related false positives?
5 comments

We appreciate the reviewer's recognition of the practical importance of VDFM and the contributions of DISTURBID-180k and our evaluation across diverse conditions. Our work highlights the tension between false alarms caused by benign disturbances and missed genuine tampering, and offers both a benchmark and a provenance-aware solution. We believe these contributions are important for reliable real-world ID document forensics. For a more coherent discussion, we first present qualitative evidence, then discuss the FPR/FNR trade-off and direct comparisons with related methods, followed by MLLM priors and the two-branch design. We therefore respond in the order W4 → Q5 → W1/Q1 → W2/Q2 → W3/Q3/Q4.

1. R3-W4: Figure 4 shows how provenance awareness and LSM preserve tampering evidence while reducing false alarms

We revised the qualitative comparison so that the differences in how each method handles VD and genuine tampering can be seen directly.

  1. We reorganized the original appendix Figure 8, included DCLNet, retained and clearly labeled MAER-only, and moved the comparison to Figure 4 in the main paper.
  2. Figure 4 includes input images, tampering GT, VD GT, and predictions from clean-only, VD-aware, MAER-only, and full CLEAR-ID. We use red for FP, yellow for FN, and green for TP, with challenging regions highlighted to make false alarms and missed tampering easier to inspect.
  3. VD-aware methods tend to suppress all anomalous responses. CLEAR-ID works differently. Through provenance-aware representation, MAER semantically distinguishes most disturbance artifacts that carry only VD semantics. It reduces false alarms while retaining more critical anomalous responses. With LSM, the model identifies key tampering factors in a low-rank subspace, allowing finer discrimination between actual tampering and disturbance-induced false positives. This explains the improved FPR/FNR trade-off, as also discussed in Section 6.3 of the revised paper.

2. R3-Q5: CLEAR-ID achieves the lowest disturbance-related FPR at the same tampering recall

The reviewer correctly notes the fixed-threshold trade-off: VD-aware TTDMamba has 0.30 UVD FPR and 38.47 FNR, while CLEAR-ID has 0.35 FPR and 31.68 FNR. This difference reflects how the two methods handle anomalous responses: suppressing them can reduce certain false alarms, but also increases missed genuine tampering.

To answer the matched-recall comparison directly, we added Figure 5 (UVD FPR–FNR trade-off curves) to the main paper. The curves show that provenance-aware representation produces a better FPR/FNR trade-off than implicit VD-aware suppression, and that LSM improves this balance further, placing CLEAR-ID on the best trade-off frontier. Thus, at the same tampering recall (the same FNR), CLEAR-ID achieves the lowest disturbance-related FPR.

3. R3-W1 / R3-Q1: Direct FPR/FNR comparisons and trade-off curves confirm CLEAR-ID's advantage under VDFM

We agree with the reviewer's suggestion. Aggregate F1 alone does not separate false alarms caused by benign VD from missed genuine tampering. A direct comparison of FPR and FNR with related methods such as DCLNet, using the same metric definitions, is therefore important.

Accordingly, we revised Table 1 to report disturbance-specific FPR, tampering-specific FNR, pixel-level F1pF1_p, and image-level F1iF1_i consistently across methods, including DCLNet, as requested. The original IoU and AUC results are still fully reported in Appendix C.1.1, Table 17. As Table 1 shows, CLEAR-ID performs better in the VDFM setting: it effectively controls false positives from benign disturbances while reducing missed genuine tampering.

We also added Figure 5, which compares UVD FPR–FNR trade-off curves with representative methods including DCLNet, FFDN, and ASCFormer. CLEAR-ID remains on the leading trade-off frontier even with these additional comparisons, further demonstrating its advantage for VDFM.

R3-W2 / R3-Q2: Controlled comparisons with the same spatial prior show the contribution of MLLM semantics

We thank the reviewer for asking us to clarify what information each prior provides. In revised Table 3 (formerly Table 2), we distinguish spatial and semantic priors and include a lightweight alternative.

(1) Controlled comparisons with the same spatial prior isolate the contribution of MLLM semantics.

All methods use the same training data, training settings, and evaluation protocol. Revised Table 3 distinguishes the information provided by each DP creator:

  • No prior → SAM (GVG_V only): isolates the effect of introducing a spatial VD prior.
  • VD-GT Masks (GVG_V only): uses the ground-truth VD mask as a spatial prior, serving as a reference upper bound for spatial information alone, not for the full semantic prior.
  • Qwen3-VL-4B / CLEAR-ID-4B: both use the same SAM spatial prior, SAM$G_V$. The comparison therefore isolates the difference between the untrained semantic prior in Qwen3-VL-4B and the MLLM semantic prior used in CLEAR-ID-4B.
  • SegFormer (SDS_D+GVG_V): provides a lightweight alternative under the same downstream setting.

Full revised Table 3 (including both UVD FPR and FNR):

VD prior / DP creatorCID F1pF1_p ↑UVD F1pF1_p ↑UVD FPR ↓UVD FNR ↓
No MLLM (No priors)69.4266.480.8533.38
SAM (GVG_V only)71.2864.770.6739.38
VD-GT Masks (GVG_V only)72.0966.100.7735.52
SegFormer (SDS_D+GVG_V)66.8363.751.0134.35
Qwen3-VL-4B (untrained SDS_D + SAM$G_V$)72.3162.220.6443.24
CLEAR-ID-4B (SDS_D + SAM$G_V$)75.6474.970.3531.68

Note: The last two rows share the SAM spatial prior (SAM$G_V$), allowing us to isolate the contribution of the MLLM semantic prior SDS_D. All values are in %; only column-best values are bold.

(2) Lightweight SegFormer cannot replace MLLM semantic priors.

We replace the MLLM-based VD-prior generator with SegFormer, keeping the other CLEAR-ID components and training/evaluation protocols unchanged. SegFormer reaches 63.75 UVD F1pF1_p, compared with 74.97 for CLEAR-ID-4B; their UVD FPR/FNR are 1.01/34.35 and 0.35/31.68, respectively. Even when SegFormer fits VD spatial localization well, its semantics do not provide equally useful provenance information.

(3) MLLM zero-shot semantics generalize to unseen VDs and can be transferred across forensic backbones.

MLLM semantic priors help distinguish benign, unseen VDs from genuine tampering. The gain magnitude varies across forensic backbones, but all evaluated experts benefit from CLEAR-ID, showing that the framework is not tied to a single expert. The same open-world semantic knowledge can be reused across different backbones, instead of requiring separate lightweight VD detectors for different disturbance types and forensic models. This gives CLEAR-ID value as a backbone-compatible provenance-aware framework, beyond the gain on any one expert.

5. R3-W3 / R3-Q3 / R3-Q4: Gains mainly come from the design and MLLM semantics, not explicit VD supervision; branch probes explain provenance awareness and modulation

To examine the contribution of auxiliary supervision, the roles of the two branches, and their behavior in overlapping regions, we added a standard Multi-Task baseline and probes on FdF_d, FtF_t, and Δt\Delta_t.

(1) The gains mainly come from our method and MLLM semantics, rather than explicit VD-task supervision.

Starting from the same TTDMamba backbone, we add a VD-specific prediction head and jointly train the model with VD and tampering supervision. Revised Table 2 (formerly Table 3) reports the comparison:

ConfigurationSFD F1p ↑UVD F1p ↑UVD FPR ↓UVD FNR ↓
TTDMamba (Clean-only)57.8355.750.9846.31
TTDMamba (Oracle-GT)53.4354.520.0053.36
TTDMamba (Multi-Task)65.8864.840.7537.72
TTDMamba (VD-aware)68.3471.000.3038.47
MAER only72.8973.680.4331.72
MAER + LSM (CLEAR-ID)74.3774.970.3531.68

Note: Values are from the revised manuscript. Only the best value in each column is bold; ↑ higher is better, ↓ lower is better. Performance values are in % unless stated otherwise.

The results show that explicit VD-task supervision does help: Multi-Task increases UVD F1pF1_p from 55.75 (clean-only) to 64.84. However, it remains below 71.00 for VD-aware TTDMamba, and clearly below 73.68 for MAER-only and 74.97 for full CLEAR-ID. Therefore, simply adding a supervised VD task does not account for our gains. The main contributions come from MAER's provenance-aware representation guided by MLLM semantics and LSM's selective modulation of tampering features.

(2) Relative metrics show that FdF_d is more VD-oriented, while FtF_t is relatively more tampering-oriented.

We evaluate both MAER branches using VD and tampering probes, and introduce two relative metrics against the original expert representation FeF_e: ΔV\Delta_V for VD and ΔT\Delta_T for tampering as shown in Table 25 .

RepresentationVD task F1p ↑Tampering task F1p ↑VD task ΔV\Delta_VTampering task ΔT\Delta_T
FeF_e34.3862.260.000.00
FdF_d64.5461.09+30.16-1.17
FtF_t63.8661.58+29.48-0.68
Δt\Delta_t40.2663.69+5.88+1.43

Note: Values are from the revised manuscript. Only the best value in each column is bold; ↑ higher is better, ↓ lower is better. Performance values are in % unless stated otherwise.

Both branches substantially improve VD perception, but FdF_d has a larger ΔV\Delta_V (+30.16 vs. +29.48), indicating a stronger VD-oriented preference. On tampering, FtF_t shows a smaller decrease in ΔT\Delta_T (-0.68 vs. -1.17), indicating better preservation of tampering evidence. Thus, both branches capture VD provenance information but develop different functional preferences.

This mechanism is also reflected in the FPR/FNR trade-off in Table 2: on UVD, Multi-Task records 0.75/37.72, MAER-only improves this to 0.43/31.72, and full MAER+LSM further reaches 0.35/31.68. In other words, MAER primarily perceives and distinguishes the provenance of VD through semantic information while preserving important anomalous responses; LSM then exploits tampering-oriented factors in the subspace to achieve a finer balance between disturbance-induced false alarms and missed genuine tampering. The probe results and the FPR/FNR ablation consistently support this interpretation.

(3) What happens where VD overlaps tampering?

During data construction we tried to avoid large overlaps, but we also analyzed the overlapping case. Unlike models that suppress all anomalous responses, MAER's FdF_d tends to preserve overlapping regions containing real tampering. This may create some additional false alarms around the boundary, but it substantially reduces missed tampering. LSM then uses selective modulation in a low-rank space to reduce these false alarms and excessive suppression, achieving a better FPR–FNR balance. The third row of Figure 4 illustrates this behavior.

Finally, we sincerely hope that:

  • Reconsideration: We are encouraged that another reviewer has recommended acceptance, recognizing the contributions of our work and the effectiveness of our rebuttal. If our responses have adequately addressed your concerns, we would greatly appreciate it if you could reconsider and, if appropriate, increase your evaluation score.
  • Further clarification: If any concerns remain, please do not hesitate to let us know. We would be more than happy to provide further clarification or make additional revisions to address them.

Thank you once again for your time, recognition, and valuable guidance!

Reviewer_1eGcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies Visual Disturbance-Induced Forensic Misattribution (VDFM), where benign visual disturbances in ID documents are mistaken for tampering. The authors introduce DISTURBID-180k, a dataset with 42 disturbance types, and propose CLEAR-ID, which uses MLLM-based disturbance priors to distinguish disturbance and tampering evidence. The method combines adaptive evidence representation (MAER) with low-rank selective modulation (LSM). Experiments show improvements in tampering detection and localization across different disturbance settings.

Major Comment

This paper addresses a relevant problem, introduces a useful dataset, and demonstrates clear performance improvements. Although the necessity of the MLLM and the generalization claims need further justification, the overall contribution is solid and I lean toward acceptance.

Critical strengths

1. Meaningful problem and dataset. The paper addresses a practical problem that existing ID document forensic methods often overlook. DISTURBID-180k provides a useful benchmark with diverse disturbances and separate annotations for disturbances and tampering.

2. Solid experimental results. CLEAR-ID consistently improves over the TTDMamba baseline, particularly under visual disturbances. The ablation studies and cross-domain experiments further support the effectiveness of the proposed method.

Critical weaknesses

1. Unclear necessity of the MLLM. The paper shows that disturbance priors are useful, but does not establish why a large MLLM is needed. A lightweight disturbance classifier or segmentation model with the same supervision would be an important baseline, especially considering the additional computational cost.

2. Limited analysis of generalization. The unseen-disturbance evaluation changes disturbance types, forgery models, and document types simultaneously, making it difficult to isolate generalization to unseen disturbances. In addition, Table 24 shows that the method can degrade performance with certain backbone initializations.

Questions for authors

Q1. Can a lightweight disturbance recognition or segmentation model trained with the same supervision achieve similar results to the MLLM-based prior?

Q2. Can the authors evaluate unseen disturbance types while keeping forgery models and document types fixed?

Q3. Why does CLEAR-ID degrade performance with certain backbones and initializations in Table 24?

4 comments

1. R2-W1 / R2-Q1: The MLLM semantic prior is necessary and cannot be replaced by a lightweight model

Thank you for your positive assessment of our work. We especially appreciate that you recognized the importance of VDFM and the value of DISTURBID-180k, including its diverse disturbances and separate VD/tampering annotations. We are also glad you highlighted CLEAR-ID's consistent improvements over TTDMamba, especially under visual disturbances, and the support from our ablation and cross-domain experiments. These strengths show why provenance-aware ID-document forensics deserves further study. We respond in the order W1/Q1 → W2/Q2 → Q3.

Thank you for asking about the necessity of the MLLM. We directly test whether a lightweight replacement can provide the same semantic benefit.

(1) A lightweight model cannot replace the VD semantic priors provided by the MLLM.

We replace CLEAR-ID's MLLM-based VD-prior component with the Transformer-based SegFormer, keeping the other modules and training/evaluation protocol unchanged. The results are added to revised Table 3 (previously Table 2):

VD prior / DP creatorCID F1pF1_p ↑UVD F1pF1_p ↑UVD FPR ↓UVD FNR ↓
No MLLM (No priors)69.4266.480.8533.38
SAM (GVG_V only)71.2864.770.6739.38
VD-GT Masks (GVG_V only)72.0966.100.7735.52
SegFormer (SDS_D+GVG_V)66.8363.751.0134.35
Qwen3-VL-4B (untrained SDS_D + SAM$G_V$)72.3162.220.6443.24
CLEAR-ID-4B (SDS_D + SAM$G_V$)75.6474.970.3531.68

Note: All values are percentages and are taken from revised Table 3. The last two rows share the SAM spatial prior (SAM$G_V$) and differ in the semantic prior. ↑ higher is better, ↓ lower is better; only the best value in each column is bold.

SegFormer learns some VD spatial information, but its semantic prior is too weak and can introduce additional noise into the forensic backbone. Its UVD F1pF1_p is 63.75 and FPR is 1.01, compared with 74.97 and 0.35 for CLEAR-ID-4B. Thus, the lightweight alternative cannot replace the MLLM semantic prior we need.

(2) Better VD spatial localization does not overcome weak semantics: SegFormer cannot even surpass the VD-GT spatial-prior reference.

An MLLM has a much richer semantic space and stronger zero-shot recognition and understanding of open-world VD types. Figure 6 compares VD-localization and forensic-task performance and shows that SegFormer can fit VD spatial localization better without achieving better final tampering detection. This agrees directly with Table 3: SegFormer reaches only 63.75 UVD F1pF1_p, below even the 66.10 obtained using VD-GT as a spatial prior. The reason is clear: its spatial prior can be strong, but its semantic prior is insufficient, so better VD localization alone cannot surpass the VD-GT spatial-prior reference. In contrast, CLEAR-ID-4B reaches 74.97 through the MLLM's richer semantic prior.

(3) The MLLM prior brings transferable gains across forensic backbones.

Revised Table 30 reports CLEAR-ID's computational cost (4,747.48M parameters, 5,035.36 GFLOPs, 119.56 ms latency, and 9,960.44 MiB peak memory). Although the gain magnitude varies across forensic backbones, all evaluated experts benefit from the provenance-aware framework. The MLLM's zero-shot semantic knowledge of open-world VDs can be reused across different experts, rather than requiring a separate lightweight VD detector for every disturbance type and forensic model. This cross-backbone reuse reduces duplicated model training, deployment, and maintenance, and is an important part of the practical value of the MLLM semantic prior.

2. R2-W2 / R2-Q2: The OCD controlled test isolates unseen VD and confirms CLEAR-ID's generalization advantage

(1) We construct OCD by keeping forgery models and document types seen during training, while changing only the VD types.

To isolate generalization to unseen visual disturbances, we sample different ID-document images from the training-data pool, using forgery models and document types already seen during training. We then render unseen VD types on these samples to build the OCD (Only Cross VD) test set. In this way, the main change comes from VD, rather than changing forgery models and document types at the same time.

For a clearer comparison, we also sample an ID (In-ID) test set from the corresponding data pool to examine how well each method fits the in-distribution data. The results are in revised Tables 19 (ID) and 20 (OCD). We place both test sets in one table and report only the four key metrics: pixel-level F1pF1_p, image-level F1iF1_i, FPR, and FNR.

MethodID F1pF1_p ↑ID F1iF1_i ↑ID FPR ↓ID FNR ↓OCD F1pF1_p ↑OCD F1iF1_i ↑OCD FPR ↓OCD FNR ↓
TruFor21.0764.381.7873.2819.5962.612.0675.40
Mesorch30.2561.090.8471.5427.5760.801.0174.04
SparseViT12.2162.410.9489.1413.7360.931.2487.00
CAT-Net35.3277.682.7536.5138.4974.183.1829.63
DTD65.8289.700.7823.8267.6685.770.8023.85
ASCFormer51.9571.770.3356.6052.4873.310.4354.99
TAFE-Net67.7588.511.166.5868.7587.551.286.47
CD-SD82.2494.490.3612.2879.8790.340.5013.10
DCLNet30.9055.151.2565.3934.0355.901.6858.37
FFDN73.1197.320.809.4674.4393.670.8211.02
TTDMamba81.4396.990.1524.0679.5693.440.2025.81
TTDMamba + LISA94.92100.000.103.3280.8689.300.2820.40
SIDA-7B24.0066.242.9258.0317.5164.554.5763.44
CLEAR-ID87.6998.060.1414.4085.0593.740.2017.10

Note: ID and OCD results are from revised Tables 19 and 20, respectively. All values are percentages. Only the best value within each column is bold (ties share the marking). F1pF1_p and F1iF1_i: higher is better; FPR and FNR: lower is better.

CLEAR-ID still performs very well on OCD with unseen VD, reaching 85.05 F1pF1_p, 93.74 F1iF1_i, 0.20 FPR, and 17.10 FNR.

The TTDMamba + LISA cascade provides an especially useful comparison. Unlike SIDA-7B, this cascade receives external tampering-forensic information and uses it to fine-tune the large model. With its large parameter capacity, it can quickly fit our in-distribution data, achieving 94.92 F1pF1_p and 100.00 F1iF1_i on ID, even exceeding CLEAR-ID. However, when only VD types change in OCD, its F1pF1_p drops from 94.92 to 80.86, its F1iF1_i drops from 100.00 to 89.30, and its FNR rises from 3.32 to 20.40. In comparison, CLEAR-ID's F1pF1_p changes only from 87.69 to 85.05, while its FNR changes from 14.40 to 17.10. This shows that a cascade relying on external tampering information to fit the training data is less suitable for generalization across VD types, and further highlights the advantage of our VD-aware design.

3. R2-Q3: Backbone architectures explain positive versus negative VD-aware transfer: preserving fine-grained forensic evidence is crucial

Revised Table 28 reveals two connected patterns.

(1) How a backbone preserves tampering evidence explains its response to VD-aware training.

ASCFormer combines transformed-domain auxiliary streams, cross-domain fusion, and tampered/authentic contrastive learning. TAFE-Net extracts multi-frequency cues and aggregates text evidence horizontally and vertically. Both rely on aggregating transformed-domain or text-structural responses, which benign VD can also activate. When optimization starts directly from generic pretrained weights, disturbance responses may become entangled with subtle forgery cues, making further modulation harmful. In Table 27, adding CLEAR-ID under pretrain-init reduces UVD F1pF1_p from 65.16 to 55.51 for ASC-Former and 41.53 to 15.26 for TAFE-Net.

By contrast, FFDN uses zero-initialized visual enhancement to preserve RGB evidence and frequency decomposition to retain small high-frequency traces. TTDMamba combines high-frequency aggregation, selective scanning, and semantic axial attention to model local tampering cues and context. **These designs can retain more usable forensic evidence under VD. **Accordingly, the same pretrain-init comparison improves UVD F1pF1_p from 54.81 to 62.16 for FFDN and 71.00 to 74.97 for TTDMamba. The key distinction is whether discriminative tampering evidence survives VD interference.

(2) Clean-init reverses some negative transfer, showing why reliable expert evidence is essential to CLEAR-ID. On CID, ASC-Former's F1pF1_p drops from 73.06 (clean-only) to 71.40 (clean-init), and TAFE-Net's from 39.20 to 35.17. In contrast, FFDN improves from 66.88 to 73.65, and TTDMamba from 63.39 to 69.33. Thus, the original VD-aware adaptation already hurts the former two backbones but not the latter two. Starting from clean-trained forensic weights and then applying CLEAR-ID restores and surpasses clean-only performance: on UVD, ASC-Former rises from 58.71 to 72.14 and TAFE-Net from 38.73 to 67.77. This demonstrates that selective modulation needs an expert to supply discriminative fine-grained forgery cues first; without such evidence within the fixed training budget, pretrain-init can cause negative transfer. Clean-init also better matches practical transfer, where an expert trained on clean ID images is adapted to VD conditions.

I appreciate the author’s detailed rebuttal, which has largely addressed my concerns. I have decided to maintain my original score and recommend acceptance.

Reviewer_bH2JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CLEAR-ID, an MLLM-guided framework for ID document forgery detection and localization under visual disturbances (VD). It combines MLLM-derived VD priors with forensic expert features through adaptive evidence representation and low-rank selective modulation. The authors also introduce DISTURBID-180k, a dataset covering 42 VD types. Experiments report improved forgery detection and localization under both seen and unseen visual disturbances.

Critical strengths

(1) The paper addresses visual disturbance-induced forensic misattribution in ID document forgery detection, a practically relevant yet underexplored problem.

(2) The authors introduce DISTURBID-180k, a large-scale dataset covering 42 types of visual disturbances, providing a useful benchmark for evaluating forensic robustness.

(3) The manuscript is well organized and clearly structured.

Critical weaknesses

(1) The manuscript contains numerous typos and grammatical errors, such as “primary” in Line 118, the incomplete sentence beginning with “Arising” in Lines 135–136, and “use” in Line 367.

(2) Highly questionable experimental results. The performance of TruFor, Mesorch, SparseViT, DTD, and SIDA-7B in Table 1 is abnormally low. These results go far beyond ordinary performance differences and raise serious concerns about the fairness and reliability of the experimental comparisons.

(3) The necessity of using an MLLM is insufficiently justified. While the authors argue that MLLMs provide useful semantic priors, they do not demonstrate why simpler, lightweight vision models cannot serve the same purpose. According to Table 26, CLEAR-ID incurs approximately 6×6\times higher inference latency and 19×19\times higher peak memory usage than TTDMamba, despite relatively modest performance gains. The authors should justify whether this substantial computational overhead is necessary and provide comparisons with lightweight alternatives.

Questions for authors

(1) Please correct these typos and grammatical errors.

(2) The authors should provide a convincing explanation for these unusually poor results and clarify how fair comparisons were ensured.

(3) The authors should justify the necessity of using an MLLM despite its substantial computational overhead and explain why simpler, lightweight alternatives would not suffice. Comparisons with lightweight alternatives should also be provided.

3 comments

R1-W2 / R1-Q2 — Why some baselines score poorly, and how we verified the comparison

We thank the reviewer for recognizing the practical importance of VDFM, the benchmark value of DISTURBID-180k covering 42 visual disturbances, and the clarity of our presentation. We also appreciate the careful scrutiny of the unusual baseline scores, which prompted further experimental checks.

Concern: Why are several baseline scores unusually low, and is the comparison fair? We included these methods because related benchmarks also evaluate them. To investigate the causes, we added an in-domain ID test (Table 19), training/validation curves (Figures 10–11), and inference checks.

1. Curves confirm learning behavior, but expose poor validation performance. Figures 10–11 show large training-loss reductions for Mesorch/SparseViT despite low validation IoU, and slower convergence for TruFor. SIDA-7B also shows limited validation gains. Late-stage validation IoU stabilizes (meeting early-stopping conditions), pointing to model/task design rather than insufficient epochs. Unlike TTDMamba + LISA, SIDA-7B has no external forensic input or explicit tampered-text annotations, making tiny character-level traces harder to learn.

2. Domain, resolution, and forgery-scale shifts explain the design mismatch. DOCFORGE-BENCH [1] reports 10–30% tampered pixels in natural images versus 0.27–4.17% in documents; our Table 8 records 1.88% clean-tampered and 0.19% VD-tampered pixels. TruFor’s noise/statistical cues face distribution shift; Mesorch’s multiscale features are less suited to tiny strokes at limited resolution; SparseViT emphasizes low-level anomalies also triggered by benign VD. All three may miss character-level evidence.

3. ID results show that several methods struggle even in-domain. Full UVD results appear in revised Table 1.

MethodID F1pF1_p ↑ID F1iF1_i ↑ID FPR ↓ID FNR ↓
TruFor21.0764.381.7873.28
Mesorch30.2561.090.8471.54
SparseViT12.2162.410.9489.14
DTD (corrected)65.8289.700.7823.82
FFDN73.1197.320.809.46
TTDMamba81.4396.990.1524.06
SIDA-7B24.0066.242.9258.03
TTDMamba + LISA94.92100.000.103.32
CLEAR-ID87.6998.060.1414.40

TruFor/Mesorch/SparseViT/SIDA-7B reach only 21.07/30.25/12.21/24.00 ID F1pF1_p: their weak evidence extraction already appears in-domain. By contrast, TTDMamba + LISA drops from 94.92 (ID) to 32.39 (UVD): in-domain fitting does not ensure VDFM generalization. We are analyzing failure cases and will add the findings.

4. DTD had a real inference error, which we corrected. We thank the reviewer: its validation curve contradicted the original test scores. Training DCT extraction was accurate and consistent; DTD was retrained. Old inference resized images to 512×512, recompressed to Q90 JPEG, and used a fixed Q90 quantization table. The wrong DCT coefficients disabled its frequency branch. The correct inference procedure uses original resolution, 512×512 sliding windows, source-JPEG DCT coefficients, and the source quantization table. We rechecked all baselines’ inference and updated Table 1 (DTD F1p/F1iF1_p/F1_i): CID 36.46/92.01; SFD 35.25/90.14; UVD 35.94/92.69. CLEAR-ID still reaches 74.97 UVD F1pF1_p.

5. The comparison is auditable, but we recognize task differences. We use common data splits, metrics, and 100 epochs, while retaining method-specific preprocessing/losses (Appendix B.1–B.2, Tables 14–16), and plan to release distributable checkpoints/materials. If a baseline’s original task makes its comparison uninformative, we are willing to remove it.

[1] Zhao et al., DOCFORGE-BENCH: A Comprehensive 0-shot Benchmark for Document Forgery Detection and Analysis, arXiv:2603.01433, 2026.

2. R1-W3 / R1-Q3: A lightweight model cannot replace the MLLM semantic prior; the prior can be transferred across forensic backbones

Thank you for raising this important question. We added a lightweight replacement experiment and report CLEAR-ID's computational cost, its gains across forensic backbones, and the efficiency benefits of caching VD priors.

(1) A lightweight model cannot replace the VD semantic priors provided by the MLLM.

We replace CLEAR-ID's MLLM-based VD-prior component with the Transformer-based SegFormer, keeping the other modules and training/evaluation protocol unchanged. The results are added to revised Table 3 (previously Table 2):

VD prior / DP creatorCID F1pF1_p ↑UVD F1pF1_p ↑UVD FPR ↓UVD FNR ↓
No MLLM (No priors)69.4266.480.8533.38
SAM (GVG_V only)71.2864.770.6739.38
VD-GT Masks (GVG_V only)72.0966.100.7735.52
SegFormer (SDS_D+GVG_V)66.8363.751.0134.35
Qwen3-VL-4B (untrained SDS_D + SAM GVG_V)72.3162.220.6443.24
CLEAR-ID-4B (SDS_D + SAM GVG_V)75.6474.970.3531.68

Note: All values are percentages and are taken from revised Table 3. The last two rows share the SAM spatial prior (SAM GVG_V) and differ in the semantic prior. ↑ higher is better, ↓ lower is better; only the best value in each column is bold.

SegFormer learns some VD spatial information, but its semantic prior is too weak and can introduce additional noise into the forensic backbone. Its UVD F1pF1_p is 63.75 and FPR is 1.01, compared with 74.97 and 0.35 for CLEAR-ID-4B. Thus, the lightweight alternative cannot replace the MLLM semantic prior required by CLEAR-ID.

(2) Better VD spatial localization does not overcome weak semantics: SegFormer cannot even surpass the VD-GT spatial-prior reference.

An MLLM has a much richer semantic space and stronger zero-shot recognition and understanding of open-world VD types. Figure 6 compares VD-localization and forensic-task performance and shows that SegFormer can fit VD spatial localization better without achieving better final tampering detection. This agrees with Table 3: SegFormer reaches only 63.75 UVD F1pF1_p, below even the 66.10 obtained using VD-GT as a spatial prior. Its spatial prior can be strong, but its semantic prior is insufficient; better VD localization alone therefore cannot surpass the VD-GT spatial-prior reference. In contrast, CLEAR-ID-4B reaches 74.97 through the MLLM's richer semantic prior. This demonstrates why MLLM-level semantic understanding is important for VDFM.

(3) The added cost brings gains across forensic experts, and the MLLM prior can be transferred across backbones.

Revised Table 30 reports CLEAR-ID's computational cost: 4,747.48M parameters, 5,035.36 GFLOPs, 119.56 ms latency, and 9,960.44 MiB peak memory. Although the gain magnitude varies, improvements are observed across the evaluated forensic experts. Thus, the added cost does not serve only one expert: CLEAR-ID is a backbone-compatible provenance-aware framework that enhances different forensic models. Its MLLM-based zero-shot semantic knowledge of open-world VDs can be reused across experts. In practice, relying only on lightweight VD detectors may require separate models for different VD types, with additional training, deployment, and maintenance. The reusable semantic prior therefore provides value across forensic experts, rather than a one-off gain on a single backbone.

(4) Caching and reusing VD priors reduces redundant inference cost.

The cost of generating VD priors can be separated from subsequent forensic inference. As reported in Table 30, the CLEAR-ID forensic expert requires 34.89 ms per image (28.91 FPS) once the VD priors are available, compared with 119.56 ms for the full pipeline. After generation, the image-specific semantic and spatial priors can be cached and reused without invoking the MLLM-based VD detector again. With the priors already available, forensic inference latency is approximately 70.8% lower than that of the full pipeline.

This enables two forms of reuse. First, the priors of the same image can be reused across repeated forensic inference or iterative analysis, avoiding redundant MLLM inference. Second, the same image-specific priors can be shared across different forensic backbones, allowing multiple forensic experts to use the same semantic and spatial guidance without regenerating the priors for each backbone.

3. R1-W1 / R1-Q1: We corrected the specified grammar errors and proofread the whole paper

Thank you for pointing these out. We corrected the specific issues and systematically proofread the paper:

  1. Original line 118: “we primary focus” → “we primarily focus.”
  2. Original lines 135–136: replaced the fragment beginning with “Arising” with “They arise from digital, physical, or hybrid operations.”
  3. Original line 367: “Training use ...” → “Training uses ...”
  4. Full-paper proofreading: we checked terminology, tense, articles, singular/plural forms, punctuation, and figure/table references.

Finally, thank you again for your thoughtful review, constructive feedback, and recognition of our work. We sincerely hope that:

  • Reconsideration: We are encouraged that another reviewer has recommended acceptance, recognizing the contributions of our work and the effectiveness of our rebuttal. If our responses have adequately addressed your concerns, we would greatly appreciate it if you could reconsider and, if appropriate, increase your evaluation score.
  • Further clarification: If any concerns remain, please do not hesitate to let us know. We would be more than happy to provide further clarification or make additional revisions to address them.

Thank you once again for your time, recognition, and valuable guidance!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.